def stoa_track(idx,
frame_counter,
img,
gt_bbox,
tracker1,
template_dir=None,
img_names=None):
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
lost_number = 0
if idx == frame_counter:
init_gt = gt_bbox_
if template_dir is not None:
img = cv2.imread(template_dir)
tracker1.init(img, gt_bbox_)
# pred_bboxes.append(1)
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
append = 1
else:
append = gt_bbox_
elif idx > frame_counter:
if img_names is not None:
img = cv2.imread(img_names[idx])
outputs = tracker1.track(img, idx=idx)
# print('****************** state of the art tracking ******************')
append = outputs['bbox']
overlap = vot_overlap(outputs['bbox'], gt_bbox,
(img.shape[1], img.shape[0]))
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
if overlap > 0:
# not lost
lost = False
else:
# lost object
append = 2
frame_counter = idx + 5 # skip 5 frames
lost_number = 1
lost = True
else:
if overlap <= 0:
lost_number = 1
else:
append = 0
return append, lost_number, frame_counter
def stoa_track(idx, frame_counter, img, gt_bbox, tracker1):
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
lost_number = 0
if idx == frame_counter:
init_gt = gt_bbox_
tracker1.init(img, gt_bbox_)
# pred_bboxes.append(1)
append = 1
elif idx > frame_counter:
outputs = tracker1.track(img, idx=idx)
# print('****************** state of the art tracking ******************')
append = outputs['bbox']
overlap = vot_overlap(outputs['bbox'], gt_bbox,
(img.shape[1], img.shape[0]))
if args.dataset != 'OTB100':
if overlap > 0:
# not lost
lost = False
else:
# lost object
append = 2
frame_counter = idx + 5 # skip 5 frames
lost_number = 1
lost = True
else:
if overlap <= 0:
lost_number = 1
else:
append = 0
return append, lost_number, frame_counter
def test_update(cfg, dirmanager):
print(cfg["TEST"]["DATASET"] + ' stage' + str(cfg["TEMPLATE"]["STEP"]))
# load config
# if cfg["TEST"]["DATASET"] == 'UAV123':
# dataset_root = '/home/lyuyu/dataset/UAV123/data_seq/UAV123/'
# else:
dataset_root = '/home/lyuyu/dataset/' + cfg["TEST"]["DATASET"]
# tracker
model_path = cfg["MODEL"]["CHECKPOINT_PATH"]
torch.cuda.set_device(cfg["TEST"]["GPU_ID"])
# load tracker and updatenet
tracker = tracker_builder.build_tracker(cfg)
# update_path='./updatenet/checkpoint/checkpoint40.pth.tar'
update_path = cfg["UPDATE"]["CHECKPOINT_PATH"]
step = cfg["TEST"]["TYPE"]
gpu_id = cfg["TEST"]["GPU_ID"]
if cfg["UPDATE"]["MODEL"][:8] == "AAUNetv2":
tracker = SiamTrackerAAUNetv2(cfg, tracker, update_path, gpu_id, step)
elif cfg["UPDATE"]["MODEL"] == "UpdateNet":
tracker = SiamTrackerUpdateNet(
cfg, tracker, update_path, gpu_id,
step) #1=dasiamrpn; 2 linear; 3 updatenet
else:
raise NotImplementedError
# create dataset
dataset = DatasetFactory.create_dataset(name=cfg["TEST"]["DATASET"],
dataset_root=dataset_root,
load_img=False)
# model_name = tracker.name
model_name = update_path[63:-7].replace('/', '').replace('.', '')
if step == 4:
model_name = 'updatenet2016'
elif step == 1:
model_name = 'dasiamrpn'
if cfg["TEST"]["DATASET"] in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
total_lost = 0
#for v_idx, video in enumerate(dataset):
if cfg["TEST"]["CLS_TYPE"] != 0:
total_success_list = []
total_iou_list = []
for video in tqdm(dataset):
# if args.video != '':
# # test one special video
# if video.name != args.video:
# continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
if cfg["TEST"]["CLS_TYPE"] != 0:
iou_list = []
success_list = []
state = dict()
for idx, (img, gt_bbox) in enumerate(video):
# print(idx)
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
state = tracker.init(img, np.array(gt_bbox))
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
pred_bbox = [cx - w / 2, cy - h / 2, w, h]
pred_bboxes.append(1)
if cfg["TEST"]["CLS_TYPE"] != 0:
iou_list.append(1)
success_list.append(1)
elif idx > frame_counter:
# state = tracker.update(img, np.array(gt_bbox))
state = tracker.update(img)
pos = state['target_pos'] # cx, cy
sz = state['target_sz'] # w, h
pred_bbox = np.array(
[pos[0] - sz[0] / 2, pos[1] - sz[1] / 2, sz[0], sz[1]])
#pred_bbox=np.array([pos[0]+1-(sz[0]-1)/2, pos[1]+1-(sz[1]-1)/2, sz[0], sz[1]])
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
# iou = overlap_ratio(gt_bbox, pred_bbox)
if cfg["TEST"]["CLS_TYPE"] != 0:
if cfg["TEST"]["CLS_TYPE"] == 1:
if overlap > cfg["UPDATE"]["IOU_THRES"]:
iou = 1
else:
iou = 0
iou_list.append(iou)
success_list.append(state['success'])
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
if cfg["TEST"]["CLS_TYPE"] != 0:
iou_list.append(0)
success_list.append(0)
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if cfg["TEST"]["VISUALIZATION"] and idx > frame_counter:
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
if cfg["TEST"]["CLS_TYPE"] != 0:
total_success_list = total_success_list + success_list
total_iou_list = total_iou_list + iou_list
success_list = np.array(success_list)
iou_list = np.array(iou_list)
# total accuracy & detect failure accuracy
accuracy = np.mean(success_list == iou_list)
index0 = np.argwhere(iou_list == 0)
accuracy0 = np.mean(success_list[index0] == iou_list[index0])
print(video.name, accuracy, accuracy0)
toc /= cv2.getTickFrequency()
# save results
if cfg["SOLVER"]["LR_POLICY"] == 'epochwise_step_group':
lr_type = cfg["UPDATE"]["CHECKPOINT_PATH"].split('/')[-2]
elif cfg["SOLVER"]["LR_POLICY"] == 'cosine':
lr_type = 'cosine'
else:
lr_type = 'undefined'
if cfg["TEST"]["TYPE"] == 1:
lr_type = 'base_dasiamrpn'
video_path = os.path.join(dirmanager.updmod_res_dir,
cfg["TEST"]["DATASET"], lr_type,
model_name, 'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
# print('({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'.format(
# v_idx+1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
if cfg["TEST"]["CLS_TYPE"] == 1:
total_success_list = np.array(total_success_list)
total_iou_list = np.array(total_iou_list)
# total accuracy & detect failure accuracy
accuracy = np.mean(total_success_list == total_iou_list)
index0 = np.argwhere(total_iou_list == 0)
accuracy0 = np.mean(
total_success_list[index0] == total_iou_list[index0])
print('total accuracy', accuracy, accuracy0)
# print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
#for v_idx, video in enumerate(dataset):
if cfg["TEST"]["CLS_TYPE"] != 0:
total_success_list = []
total_iou_list = []
for video in tqdm(dataset):
# if args.video != '':
# # test one special video
# if video.name != args.video:
# continue
toc = 0
pred_bboxes = []
if cfg["TEST"]["CLS_TYPE"] != 0:
iou_list = []
success_list = []
scores = []
track_times = []
state = dict()
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
state = tracker.init(
img, np.array(gt_bbox)) #注意gt_bbox和gt_bbox_的区别
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
pred_bbox = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
scores.append(None)
if 'VOT2018-LT' == cfg["TEST"]["DATASET"]:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
if cfg["TEST"]["CLS_TYPE"] != 0:
iou_list.append(1)
success_list.append(1)
# if video.name == 'Jogging-1':
# template_vis(state['z_f_cur'], 0, 'template_vis_'+str(idx))
else:
state = tracker.update(img)
# if video.name == 'Jogging-1':
# template_vis(state['z_f_cur'], 0, 'template_vis_'+str(idx))
pos = state['target_pos']
sz = state['target_sz']
pred_bbox = np.array(
[pos[0] - sz[0] / 2, pos[1] - sz[1] / 2, sz[0], sz[1]])
pred_bboxes.append(pred_bbox)
#scores.append(outputs['best_score'])
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if cfg["TEST"]["CLS_TYPE"] != 0:
if cfg["TEST"]["CLS_TYPE"] == 1:
if overlap > 0.1:
iou = 1
else:
iou = 0
if cfg["TEST"]["CLS_TYPE"] == 2:
iou = overlap
iou_list.append(iou)
success_list.append(state['success'].cpu().numpy())
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if cfg["TEST"]["VISUALIZATION"] and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(
img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]),
(0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3]), (0, 255, 255),
3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
if cfg["TEST"]["CLS_TYPE"] != 0:
total_success_list = total_success_list + success_list
total_iou_list = total_iou_list + iou_list
success_list = np.array(success_list)
iou_list = np.array(iou_list)
if cfg["TEST"]["CLS_TYPE"] == 1:
# total accuracy & detect failure accuracy
accuracy = np.mean(success_list == iou_list)
index0 = np.argwhere(iou_list == 0)
index1 = np.argwhere(iou_list == 1)
accuracy0 = np.mean(
success_list[index0] == iou_list[index0])
accuracy1 = np.mean(
success_list[index1] == iou_list[index1])
if cfg["TEST"]["CLS_TYPE"] == 2:
# total accuracy & detect failure accuracy
comp_list = abs(success_list - iou_list) < 0.2
accuracy = np.mean(comp_list)
index0 = np.argwhere((success_list - iou_list) > 0)
index1 = np.argwhere((iou_list - success_list) > 0)
accuracy0 = np.mean(comp_list[index0])
accuracy1 = np.mean(comp_list[index1])
print(video.name, accuracy, accuracy0, accuracy1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == cfg["TEST"]["DATASET"]:
video_path = os.path.join('results', cfg["TEST"]["DATASET"],
model_name, 'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == cfg["TEST"]["DATASET"]:
video_path = os.path.join('results', cfg["TEST"]["DATASET"],
model_name, video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
lr_type = 'cosine'
video_path = os.path.join(dirmanager.updmod_res_dir,
cfg["TEST"]["DATASET"], lr_type,
model_name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
class_path = os.path.join(video_path,
'{}_cls.txt'.format(video.name))
with open(class_path, 'w') as f:
for x in success_list:
f.write(str(x) + '\n')
# print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.format(
# v_idx+1, video.name, toc, idx / toc))
total_success_list = np.array(total_success_list)
total_iou_list = np.array(total_iou_list)
# total accuracy & detect failure accuracy
if cfg["TEST"]["CLS_TYPE"] == 1:
accuracy = np.mean(total_success_list == total_iou_list)
index0 = np.argwhere(total_iou_list == 0)
accuracy0 = np.mean(
total_success_list[index0] == total_iou_list[index0])
if cfg["TEST"]["CLS_TYPE"] == 2:
comp_list = abs(total_success_list - total_iou_list) < 0.2
accuracy = np.mean(comp_list)
index0 = np.argwhere((total_success_list - total_iou_list) > 0)
index1 = np.argwhere((total_iou_list - total_success_list) > 0)
accuracy0 = np.mean(comp_list[index0])
accuracy1 = np.mean(comp_list[index1])
print('total accuracy', accuracy, accuracy0)
# evaluation(cfg["TEST"]["DATASET"], model_name, dirmanager.updmod_res_dir)
evaluation(
cfg["TEST"]["DATASET"], model_name,
os.path.join(dirmanager.updmod_res_dir, cfg["TEST"]["DATASET"],
lr_type))
return
def track(video, tracker, visualize=False, data_collector=None):
num_frames = len(video)
frame_counter = 0
frame_reset = 0 # used to indicate how many times the tracker was re-initialized
lost_times = 0
pred_bboxes = [
] # Filled according to VOT protocol & used for metric calculation
total_time = 0
zero_tensor = torch.zeros(cfg.REFINE_TEMPLATE.FEATURE_SIZE,
dtype=torch.float32).cpu().data
for f, (im, gt) in enumerate(video):
if len(gt) == 4:
gt = bbox_to_polygon(gt)
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt))
start_time = cv2.getTickCount()
if f == frame_counter: # Init or reset after lost frame
gt_bbox = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(im, gt_bbox)
pred_bbox = gt_bbox
pred_bboxes.append(1)
total_time += cv2.getTickCount() - start_time
frame_reset = 0
if data_collector:
data_collector.add_init(f,
num_frames,
init_feat=tracker.zf_init)
elif f > frame_counter: # Tracking
frame_reset += 1
outputs = tracker.track(im)
pred_bbox = outputs['bbox']
if data_collector:
# Extract ground-truth template features
gt_rect = np.array([cx, cy, w, h])
gt_zf = tracker.extract_template(
im, gt_rect) if w * h != 0 else zero_tensor
data_collector.add_tracking(f,
num_frames,
frame_reset,
cur_feat=outputs['zf'],
pre_feat=tracker.zf,
gt_feat=gt_zf)
overlap = vot_overlap(pred_bbox, gt, (im.shape[1], im.shape[0]))
if overlap > 0:
pred_bboxes.append(pred_bbox)
else:
pred_bboxes.append(2)
# skip 5 frames after object lost (as suggested by VOT)
frame_counter = f + 5
lost_times += 1
total_time += cv2.getTickCount() - start_time
elif f < frame_counter or w * h == 0: # Skipping
pred_bboxes.append(0)
total_time += cv2.getTickCount() - start_time
frame_reset = 0
if data_collector:
data_collector.add_init(f, num_frames, zero_tensor)
if visualize:
cv2.polylines(im, [np.array(gt, np.int).reshape((-1, 1, 2))], True,
(0, 255, 0), 3)
bbox = list(map(int, pred_bbox))
cv2.rectangle(im, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(im, str(f), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1,
(0, 255, 255), 2)
cv2.putText(im, str(lost_times), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
window_name = 'test'
cv2.imshow(window_name, im)
cv2.moveWindow(window_name, 100, 10)
cv2.waitKey(1)
total_time /= cv2.getTickFrequency()
cv2.destroyAllWindows()
return {
'pred_bboxes': pred_bboxes,
'lost_times': lost_times,
'total_time': total_time,
'fps': f / total_time,
}
def main():
is_gpu_cuda_available = torch.cuda.is_available()
if not is_gpu_cuda_available:
raise RuntimeError(
'Failed to locate a CUDA GPU. Program cannot continue..')
num_gpus = torch.cuda.device_count()
gpu_type = torch.cuda.get_device_name(0)
print(f"You have {num_gpus} available of type: {gpu_type}")
print("This might take a few minutes...Grab a cup of coffee\n")
# load config
cfg.merge_from_file(args.config)
dataset_root = os.path.join(args.dataset_directory, args.dataset)
print(f"dataset root-->{dataset_root}")
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.model_name
print(f"Model name is {model_name}")
total_lost = 0
if args.dataset in vot_like_dataset:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0.85:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + args.skip_frames # skip 1 frame
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
save_path = os.path.join(args.results_path, args.dataset,
model_name, args.experiment_name,
video.name)
if not os.path.isdir(save_path):
os.makedirs(save_path)
result_path = os.path.join(save_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
with open(os.path.join(save_path, '..', 'lost.txt'), 'a+') as f:
f.write(
f"{v_idx+1} Class: {video.name} | Time: {toc}s | Speed: {idx/toc}fps | Lost:{lost_number} \n"
)
print(
'({:3d}) Class: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
with open(os.path.join(save_path, '..', 'lost.txt'), 'a+') as f:
f.write(
f"Model architeture used --> {model_name} \ntotal lost: {total_lost} \n"
)
f.write(f"SKIP FRAMES USED --> {args.skip_frames}")
else:
# OPE tracking
# will be implemented if needed in future
pass
def main():
# load config
dataset_root='/home/ubuntu/pytorch/pytorch-tracking/DaSiamRPN/datasets/'+args.dataset
# tracker
model_path ='./models/SiamRPNBIG.model'
name='DaSiamRPN'
gpu_id=0 #这里改成1不能正常运行
tracker = SiamRPNTracker(model_path,gpu_id)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
#算法的名字
model_name = name
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
total_lost=0
#for v_idx, video in enumerate(dataset):
for video in tqdm(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
state=dict()
for idx, (img, gt_bbox) in enumerate(video):
# print(idx)
if len(gt_bbox) == 4:
gt_bbox = [gt_bbox[0], gt_bbox[1],
gt_bbox[0], gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]]
tic = cv2.getTickCount()
if idx == frame_counter:
state=tracker.init(img, np.array(gt_bbox))#注意gt_bbox和gt_bbox_的区别
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox)) #1-based
pred_bbox = [cx-(w)/2, cy-(h)/2, w, h]#1-based
pred_bboxes.append(1)
elif idx > frame_counter:
state = tracker.update(img)
pos=state['target_pos']
sz=state['target_sz']
pred_bbox=np.array([pos[0]-sz[0]/2, pos[1]-sz[1]/2, sz[0], sz[1]])
#pred_bbox=np.array([pos[0]+1-(sz[0]-1)/2, pos[1]+1-(sz[1]-1)/2, sz[0], sz[1]])
overlap = vot_overlap(pred_bbox, gt_bbox, (img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(img, [np.array(gt_bbox, np.int).reshape((-1, 1, 2))],
True, (0, 255, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(img, [np.array(pred_bbox, np.int).reshape((-1, 1, 2))],
True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0]+bbox[2], bbox[1]+bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i) for i in x])+'\n')
# print('({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'.format(
# v_idx+1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
# print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
#for v_idx, video in enumerate(dataset):
for video in tqdm(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
state=dict()
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
state=tracker.init(img, np.array(gt_bbox))#注意gt_bbox和gt_bbox_的区别
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
pred_bbox = [cx-(w-1)/2, cy-(h-1)/2, w, h]
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
state = tracker.update(img)
pos=state['target_pos']
sz=state['target_sz']
pred_bbox=np.array([pos[0]-sz[0]/2, pos[1]-sz[1]/2, sz[0], sz[1]])
pred_bboxes.append(pred_bbox)
#scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append((cv2.getTickCount() - tic)/cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0]+gt_bbox[2], gt_bbox[1]+gt_bbox[3]), (0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0]+pred_bbox[2], pred_bbox[1]+pred_bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write("{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name, video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('results', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path, '{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
def main():
# load config
cfg.merge_from_file(args.config)
cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(args.dataset_dir, args.dataset)
epsilon = args.epsilon
# create model
track_model1 = ModelBuilder()
track_model2 = ModelBuilder()
lr = args.lr
# load model
track_model1 = load_pretrain(track_model1, args.snapshot).cuda().eval()
track_model2 = load_pretrain(track_model2, args.snapshot).cuda().eval()
# build tracker
tracker1 = build_tracker(track_model1)
tracker2 = build_tracker(track_model2)
attacker = ModelAttacker().cuda().train()
optimizer = optim.Adam(attacker.parameters(), lr=lr)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False,
dataset_toolkit='oneshot',
config=cfg)
#
# vid.name = {'ants1','ants3',....}
# img, bbox, cls, delta, delta_weight
# vid[0][0],vid[0][1],vid[0][2],vid[0][3],vid[0][4]
model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
n_epochs = args.epochs
for name, param in tracker1.model.named_parameters():
param.requires_grad_(False)
for name, param in tracker2.model.named_parameters():
param.requires_grad_(False)
# for name, param in tracker2.model.named_parameters():
# if 'backbone' in name or 'neck' in name or 'rpn_head' in name:
# param.requires_grad_(False)
# elif param.requires_grad:
# param.requires_grad_(True)
# # print(name, param.data)
# print('grad true ', name)
# else:
# print('grad false ', name)
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019', 'OTB100']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
else:
if not os.path.exists(
os.path.join(args.savedir, video.name)):
os.mkdir(os.path.join(args.savedir, video.name))
# set writing video parameters
height, width, channels = video[0][0].shape
out = cv2.VideoWriter(
os.path.join(args.savedir, video.name + '.avi'),
cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'), 15,
(width, height))
frame_counter = 0
frame_counter_adv = 0
lost_number = 0
lost_number_adv = 0
toc = 0
total_toc = 0
pred_bboxes = []
pred_bboxes_adv = []
lost = False
lost_adv = False
for i in range(0, args.epochs):
for idx, (img, gt_bbox) in enumerate(video):
# if len(gt_bbox) == 4:
# gt_bbox = [gt_bbox[0], gt_bbox[1],
# gt_bbox[0], gt_bbox[1] + gt_bbox[3] - 1,
# gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1] + gt_bbox[3] - 1,
# gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]]
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
##########################################
# # # for state of the art tracking # # #
##########################################
if i == 0:
if idx == frame_counter:
init_gt = gt_bbox_
tracker1.init(img, gt_bbox_)
pred_bboxes.append(1)
zf = tracker1.zf
img1 = img.copy()
elif idx > frame_counter:
outputs = tracker1.track(img, idx=idx)
# print('****************** state of the art tracking ******************')
pred_bbox = outputs['bbox']
if args.dataset != 'OTB100':
overlap = vot_overlap(
pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
lost = False
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
lost = True
else:
pred_bboxes.append(pred_bbox)
else:
pred_bboxes.append(0)
tic = cv2.getTickCount()
##########################################
# # # # # adversarial tracking # # # # #
##########################################
if idx == frame_counter_adv:
zimg = img.copy()
sz, bbox, pad = tracker2.init(img,
gt_bbox_,
attacker=attacker,
epsilon=args.epsilon)
pred_bboxes_adv.append(1)
zf2 = tracker2.zf
# cv2.imwrite(os.path.join(args.savedir, video.name, str(idx).zfill(6) +'.jpg'), img)
elif idx > frame_counter_adv:
_outputs = tracker2.track(img,
attacker=attacker,
epsilon=args.epsilon,
zf=zf2,
idx=idx,
iter=i)
# print(_outputs['best_score'], outputs['target_score'])
# ad_bbox = _outputs['bbox']
# ad_overlap = vot_overlap(ad_bbox, gt_bbox, (img.shape[1], img.shape[0]))
# filename = os.path.join(args.savedir, video.name, str(idx).zfill(6) +'.jpg')
# save_2bb(img, filename, ad_bbox, pred_bbox, gt_bbox)
filename = os.path.join(
args.savedir, video.name,
'bb' + str(idx).zfill(6) + '.jpg')
save_2bb(img, filename, ad_bbox, pred_bbox, gt_bbox)
# _zimg = save(zimg, tracker2.z_crop_adv, sz, init_gt, pad,
# os.path.join(args.savedir, video.name, str(idx).zfill(6) + '.jpg'), save=True)
# _zimg = save(zimg, tracker2.z_crop_adv, sz, init_gt, pad, os.path.join(args.savedir, video.name, str(idx).zfill(6) +'.jpg'), save=True)
# update state
tracker2.center_pos = _outputs['center_pos']
tracker2.size = _outputs['size']
ad_overlap = vot_overlap(ad_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if args.dataset != 'OTB100':
if ad_overlap > 0:
# not lost
pred_bboxes_adv.append(ad_bbox)
lost_adv = False
else:
# lost object
pred_bboxes_adv.append(2)
frame_counter_adv = idx + 5 # skip 5 frames
lost_number_adv += 1
lost_adv = True
else:
if ad_overlap <= 0:
lost_number_adv += 1
pred_bboxes_adv.append(ad_bbox)
else:
pred_bboxes_adv.append(0)
toc += cv2.getTickCount() - tic
# if idx > frame_counter_adv and not lost_adv:
# ad_bbox = list(map(int, ad_bbox))
# cv2.rectangle(img, (ad_bbox[0], ad_bbox[1]),
# (ad_bbox[0] + ad_bbox[2], ad_bbox[1] + ad_bbox[3]), (0, 0, 255), 3)
#
# if idx > frame_counter and not lost:
# bbox = list(map(int, pred_bbox))
# cv2.rectangle(img, (bbox[0], bbox[1]),
# (bbox[0] + bbox[2], bbox[1] + bbox[3]), (0, 255, 255), 3)
#
# __gt_bbox = list(map(int, gt_bbox_))
# cv2.rectangle(img, (__gt_bbox[0], __gt_bbox[1]),
# (__gt_bbox[0]+__gt_bbox[2], __gt_bbox[1]+__gt_bbox[3]), (0, 0, 0), 3)
#
# cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 0), 2)
# cv2.putText(img, str(lost_number), (40, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
# cv2.putText(img, ","+str(lost_number_adv), (80, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
#
# out.write(img)
# print('frame {}/{} Lost: {:d}'.format(idx, len(video), lost_number_adv))
toc /= cv2.getTickFrequency()
print(
'({:3d}) Video: {:12s} train{}/{} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, i, args.epochs, toc,
idx / toc, lost_number_adv))
total_toc += toc
l1 = _outputs['l1']
l2 = _outputs['l2']
l3 = _outputs['l3']
# total_loss = 0.8 * l1 + 0.4 * l2 + 1.2 * l3
total_loss = l1 + 0.4 * l2
# total_loss = l1 + 0.4 * l2
# print(idx, i, total_loss.item(), _outputs['center_pos'], _outputs['size'])
# if ad_overlap < 0.5:
if _outputs['best_score'] < outputs['target_score']:
total_loss_val = 0
# print(idx, i, ad_overlap)
# print(ad_bbox)
# print(pred_bbox)
# print('------------------------')
# filename = os.path.join(args.savedir, video.name, 'bb' + str(idx).zfill(6) + '.jpg')
# save_2bb(img, filename, ad_bbox, pred_bbox, gt_bbox)
# _zimg = save(zimg, tracker2.z_crop_adv, sz, init_gt, pad,
# os.path.join(args.savedir, video.name, str(idx).zfill(6) + '.jpg'), save=True)
# pdb.set_trace()
break
else:
# print(_outputs['bbox'])
optimizer.zero_grad()
total_loss.backward(retain_graph=True)
optimizer.step()
total_toc /= (i + 1)
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
print(
'({:3d}) Video: {:12s} Time Average: {:4.1f}s Speed Average: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, total_toc, idx / total_toc,
lost_number_adv))
total_lost += lost_number_adv
print("{:s} total lost: {:d}".format(model_name, total_lost))
def objective(trial):
# different params
cfg.TRACK.WINDOW_INFLUENCE = trial.suggest_uniform('window_influence',
0.050, 0.650)
cfg.TRACK.PENALTY_K = trial.suggest_uniform('penalty_k', 0.000, 0.600)
cfg.TRACK.LR = trial.suggest_uniform('scale_lr', 0.100, 0.800)
cfg.TRACK.COEE_CLASS = trial.suggest_uniform('coee_class', 0.01, 0.999)
# rebuild tracker
info = edict()
info.arch = args.arch
info.cls_type = args.cls_type
info.dataset = args.dataset
tracker = SiamRPN(info)
model_name = args.snapshot.split('/')[-1].split('.')[0]
tracker_name = os.path.join('tune_results', args.dataset, model_name, model_name + \
'_wi-{:.3f}'.format(cfg.TRACK.WINDOW_INFLUENCE) + \
'_pk-{:.3f}'.format(cfg.TRACK.PENALTY_K) + \
'_lr-{:.3f}'.format(cfg.TRACK.LR)+\
'_ce-{:.3f}'.format(cfg.TRACK.COEE_CLASS)
)
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
# if len(gt_bbox) == 4:
# gt_bbox = [gt_bbox[0], gt_bbox[1],
# gt_bbox[0], gt_bbox[1] + gt_bbox[3] - 1,
# gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1] + gt_bbox[3] - 1,
# gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
#gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
target_pos = np.array([cx, cy])
target_sz = np.array([w, h])
state = tracker.init(img, target_pos, target_sz, net)
state["arch"] = args.arch
# pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
state = tracker.track(state, img) # track
location = cxy_wh_2_rect(state['target_pos'],
state['target_sz'])
pred_bbox = location
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join(tracker_name, 'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
print(
'({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
eao = eval(dataset=dataset_eval, tracker_name=tracker_name)
info = "{:s} window_influence: {:1.17f}, penalty_k: {:1.17f}, scale_lr: {:1.17f}, EAO: {:1.3f}".format(
model_name, cfg.TRACK.WINDOW_INFLUENCE, cfg.TRACK.PENALTY_K,
cfg.TRACK.LR, eao)
logging.getLogger().info(info)
print(info)
return eao
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
target_pos = np.array([cx, cy])
target_sz = np.array([w, h])
state = tracker.init(img, target_pos, target_sz,
net) # init tracker
state["arch"] = args.arch
# tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
state = tracker.track(state, img) # track
location = cxy_wh_2_rect(state['target_pos'],
state['target_sz'])
pred_bbox = location
# outputs = tracker.track(img)
# pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(state['score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
if not os.path.isdir(tracker_name):
os.makedirs(tracker_name)
result_path = os.path.join(tracker_name,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.
format(v_idx + 1, video.name, toc, idx / toc))
auc = eval(dataset=dataset_eval, tracker_name=tracker_name)
info = "{:s} window_influence: {:1.17f}, penalty_k: {:1.17f}, scale_lr: {:1.17f}, AUC: {:1.3f}".format(
model_name, cfg.TRACK.WINDOW_INFLUENCE, cfg.TRACK.PENALTY_K,
cfg.TRACK.LR, auc)
logging.getLogger().info(info)
print(info)
return auc
def main():
# load config
cfg_from_file(args.config)
dataset_root = os.path.join('dataset', args.dataset)
# create model
net = ModelBuilder()
checkpoint = torch.load(args.model)
if 'state_dict' in checkpoint:
net.load_state_dict(checkpoint['state_dict'])
else:
net.load_state_dict(checkpoint)
net.cuda().eval()
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.save_name
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
target_pos, target_sz = np.array([cx,
cy]), np.array([w, h])
state = CGACD_init(img, target_pos, target_sz, net)
pred_bbox = cxy_wh_2_rect(state['target_pos'],
state['target_sz'])
pred_bboxes.append(1)
elif idx > frame_counter:
state = CGACD_track(state, img)
pred_bbox = cxy_wh_2_rect(state['target_pos'],
state['target_sz'])
pred_polygon = [
pred_bbox[0], pred_bbox[1],
pred_bbox[0] + pred_bbox[2], pred_bbox[1],
pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3], pred_bbox[0],
pred_bbox[1] + pred_bbox[3]
]
overlap = vot_overlap(gt_bbox, pred_polygon,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
target_pos = state['target_pos']
target_sz = state['target_sz']
cv2.rectangle(img, (int(target_pos[0] - target_sz[0] / 2),
int(target_pos[1] - target_sz[1] / 2)),
(int(target_pos[0] + target_sz[0] / 2),
int(target_pos[1] + target_sz[1] / 2)),
(0, 255, 0), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2,
cv2.LINE_AA)
cv2.imshow(video.name, img)
cv2.moveWindow(video.name, 100, 100)
key = cv2.waitKey(1)
if key == 27:
break
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('result', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
print(
'({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
if 'OTB' in args.dataset:
target_pos, target_sz = rect1_2_cxy_wh(gt_bbox)
else:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
target_pos, target_sz = np.array([cx, cy
]), np.array([w, h])
state = CGACD_init(img, target_pos, target_sz, net)
if 'OTB' in args.dataset:
pred_bbox = cxy_wh_2_rect1(state['target_pos'],
state['target_sz'])
else:
pred_bbox = cxy_wh_2_rect(state['target_pos'],
state['target_sz'])
pred_bboxes.append(pred_bbox)
else:
state = CGACD_track(state, img)
pred_bbox = cxy_wh_2_rect(state['target_pos'],
state['target_sz'])
pred_bboxes.append(pred_bbox)
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
target_pos = state['target_pos']
target_sz = state['target_sz']
cv2.rectangle(img, (int(target_pos[0] - target_sz[0] / 2),
int(target_pos[1] - target_sz[1] / 2)),
(int(target_pos[0] + target_sz[0] / 2),
int(target_pos[1] + target_sz[1] / 2)),
(0, 255, 0), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2,
cv2.LINE_AA)
cv2.imshow(video.name, img)
cv2.moveWindow(video.name, 100, 100)
key = cv2.waitKey(1)
if key == 27:
break
toc /= cv2.getTickFrequency()
if 'GOT-10k' == args.dataset:
video_path = os.path.join('result', args.dataset, model_name,
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('result', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.
format(v_idx + 1, video.name, toc, idx / toc))
def main():
# load config
cfg.merge_from_file(args.config)
cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(args.dataset_dir, args.dataset)
epsilon = args.epsilon
# create model
model = Steath([1, 3, 255, 255])
track_model = ModelBuilder()
lr = args.lr
# load model
model = load_pretrain(model, args.snapshot).cuda()
track_model = load_pretrain(track_model, args.snapshot).cuda().eval()
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.9)
model.train()
# model.dx.requires_grad_(True)
# model.backbone.eval()
# if cfg.ADJUST.ADJUST:
# model.neck.eval()
# model.rpn_head.eval()
for name, param in model.named_parameters():
if 'backbone' in name or 'neck' in name or 'rpn_head' in name:
param.requires_grad_(False)
elif param.requires_grad:
param.requires_grad_(True)
print(name, param.data)
else:
print(name)
clipper = WeightClipper(5)
# build tracker
tracker1 = build_tracker(track_model)
tracker2 = build_tracker(track_model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False,
config=cfg)
#
# vid.name = {'ants1','ants3',....}
# img, bbox, cls, delta, delta_weight
# vid[0][0],vid[0][1],vid[0][2],vid[0][3],vid[0][4]
model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
n_epochs = args.epochs
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
else:
if not os.path.exists(os.path.join(args.savedir, video.name)):
os.mkdir(os.path.join(args.savedir, video.name))
# set writing video parameters
height, width, channels = video[0][0].shape
out = cv2.VideoWriter(os.path.join(args.savedir, video.name + '.avi'),
cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'), 15, (width, height))
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
data = {'template': None, 'search': None}
for idx, (img, gt_bbox, z, x, szx, boxx, padx, cls, delta, delta_w, overlap, _bbox, _bbox_p) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [gt_bbox[0], gt_bbox[1],
gt_bbox[0], gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]]
tic = cv2.getTickCount()
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx-w//2, cy-h//2, w, h]
if idx == frame_counter:
tracker1.init(img, gt_bbox_)
tracker2.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
data['template'] = torch.autograd.Variable(z, requires_grad=True).cuda()
elif idx > frame_counter:
prim_img = np.copy(img)
data['search'] = torch.autograd.Variable(x, requires_grad=True).cuda()
data['label_cls'] = torch.Tensor(cls).type(torch.LongTensor).cuda()
data['label_loc'] = torch.Tensor(delta).type(torch.FloatTensor).cuda()
data['label_loc_weight'] = torch.Tensor(delta_w).cuda()
diff = data['search']
for epoch in range(n_epochs):
outputs = model(data, epsilon)
cls_loss = outputs['cls_loss']
# print(idx, epoch, cls_loss.item())
loc_loss = outputs['loc_loss']
total_loss = outputs['total_loss']
print('{}/{} cls={}, loc={}, total={}'.format(idx, len(video), cls_loss.item(), loc_loss.item(),
total_loss.item()))
optimizer.zero_grad()
# cls_loss.backward()
total_loss.backward()
# model.apply(clipper)
optimizer.step()
# print('loss ', loss(diff, outputs['search']).item())
# diff = outputs['search']
# print(epoch, cls_loss, loc_loss, total_loss)
# print('{}/{} cls={}, loc={}, total={}'.format(idx, len(video), cls_loss.item(), loc_loss.item(),
# total_loss.item()))
perturb_data = outputs['search']
# cv2.rectangle(img, (int(cx-w/2+1), int(cy-h/2+1)), (int(cx+w/2+1), int(cy+h/2+1)), (0, 0, 0), 3)
# cv2.imwrite(os.path.join(args.savedir, video.name, 'original_' + str(idx).zfill(7) + '.jpg'), img)
# _img = perturb_data.data.cpu().numpy().squeeze().transpose([1, 2, 0])
# cv2.imwrite(os.path.join(args.savedir, 'perturb_' + str(idx) + '.jpg'), _img)
szx = int(szx)
if not np.array_equal(cfg.TRACK.INSTANCE_SIZE, szx):
perturb_data = F.interpolate(perturb_data, size=szx)
__bbox = (np.array(_bbox_p)*szx/cfg.TRACK.INSTANCE_SIZE).astype(np.int)
_img = cv2.UMat(perturb_data.data.cpu().numpy().squeeze().transpose([1, 2, 0])).get()
cv2.rectangle(_img, (__bbox[0], __bbox[1]), (__bbox[2], __bbox[3]), (0, 0, 0), 3)
cv2.imwrite(os.path.join(args.savedir, video.name, 'crop_full_' + str(idx) + '.jpg'), _img)
nh, nw, _ = _img.shape
__bbox0 = np.zeros_like(__bbox)
__bbox0[:4:2] = __bbox[:4:2] - padx[0]
__bbox0[1:4:2] = __bbox[1:4:2] - padx[2]
img[boxx[0]:boxx[1] + 1, boxx[2]:boxx[3] + 1, :] = \
_img[boxx[0]+padx[0]:boxx[1]+padx[0] + 1, 0 + padx[2]:boxx[3] - boxx[2] + padx[2] + 1, :]
# cv2.imwrite(os.path.join(args.savedir, video.name, 'perturb_full_' + str(idx) + '.jpg'), img)
# if not np.array_equal(cfg.TRACK.INSTANCE_SIZE, sz):
# perturb_data = F.interpolate(perturb_data, size=sz)
# __bbox = (np.array(_bbox)*sz/cfg.TRACK.INSTANCE_SIZE).astype(np.uint8)
#
# _img = cv2.UMat(perturb_data.data.cpu().numpy().squeeze().transpose([1, 2, 0])).get()
# cv2.rectangle(_img, (__bbox[0], __bbox[1]), (__bbox[2], __bbox[3]), (0, 0, 0), 3)
# cv2.imwrite(os.path.join(args.savedir, video.name, 'crop_full_' + str(idx) + '.jpg'), _img)
#
# nh, nw, _ = _img.shape
# img[bT:bB+1, bL:bR+1, :] = _img[pad[0]:nh - pad[1], pad[2]:nw - pad[3], :]
# cv2.imwrite(os.path.join(args.savedir, video.name, 'perturb_full_' + str(idx) + '.jpg'), img)
# nimg, sz, box, pad = tracker2.crop(img, bbox=gt_bbox_, im_name='search' + str(idx))
outputs = tracker1.track(img)
prim_outputs = tracker2.track(prim_img)
pred_bbox = outputs['bbox']
prim_box = prim_outputs['bbox']
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox, (img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
# cv2.imwrite(os.path.join(args.savedir, video.name, str(idx).zfill(7) + '.jpg'), img)
toc += cv2.getTickCount() - tic
# write ground truth bbox
cv2.polylines(img, [np.array(gt_bbox, np.int).reshape((-1, 1, 2))],
True, (255, 255, 255), 3)
if idx != frame_counter:
bbox = list(map(int, pred_bbox))
prim_bbox = list(map(int, prim_box))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]), (0, 255, 255), 3)
cv2.rectangle(img, (prim_bbox[0], prim_bbox[1]),
(prim_bbox[0] + prim_bbox[2], prim_bbox[1] + prim_bbox[3]), (0, 0, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
out.write(img)
cv2.imwrite(os.path.join(args.savedir, video.name, str(idx).zfill(7) + '.jpg'), img)
# import pdb
# pdb.set_trace()
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i) for i in x])+'\n')
print('({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'.format(
v_idx+1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx-(w-1)/2, cy-(h-1)/2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append((cv2.getTickCount() - tic)/cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0]+gt_bbox[2], gt_bbox[1]+gt_bbox[3]), (0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0]+pred_bbox[2], pred_bbox[1]+pred_bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write("{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name, video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('results', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path, '{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.format(
v_idx+1, video.name, toc, idx / toc))
def vot_evaluate(dataset, tracker):
tracker_name = args.tracker
backbone_name = args.cfg.split('/')[-1].split('_')[0]
snapshot_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
for v_idx, video in enumerate(dataset):
if args.video != '': # if test special video
if video.name != args.video:
continue
frame_count = 0
lost_number = 0
pred_bboxes = []
toc = 0
for idx, (frame, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == frame_count:
tracker.init(frame, gt_bbox) # cx,cy,w,h
pred_bboxes.append(1)
elif idx > frame_count:
track_result = tracker.track(frame)
bbox = track_result['bbox'] # cx,cy,w,h
score = track_result['score']
bbox_ = [
bbox[0] - bbox[2] / 2, bbox[1] - bbox[3] / 2, bbox[2],
bbox[3]
] # x,y,w,h
gt_bbox_ = [
gt_bbox[0] - (gt_bbox[2] - 1) / 2,
gt_bbox[1] - (gt_bbox[3] - 1) / 2, gt_bbox[2], gt_bbox[3]
]
overlap = vot_overlap(bbox_, gt_bbox_,
(frame.shape[1], frame.shape[0]))
# print('idx: {}\n pred: {}\n gt: {}\n overlap: {}\n'.format(idx, bbox_, gt_bbox_, overlap))
if overlap > 0:
pred_bboxes.append(bbox_)
else:
# print('lost idx: {}'.format(idx))
pred_bboxes.append(2)
frame_count = idx + 5
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if args.vis and idx > frame_count:
show_double_bbox(frame, bbox, score, gt_bbox, idx, lost_number)
toc /= cv2.getTickFrequency()
result_dir = os.path.join(cfg.TRACK.RESULT_DIR, args.dataset,
tracker_name, backbone_name, snapshot_name)
if not os.path.isdir(result_dir):
os.makedirs(result_dir)
result_path = '{}/{}.txt'.format(result_dir, video.name)
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write('{:d}\n'.format(x))
else:
f.write(','.join(['{:.4f}'.format(i) for i in x]) + '\n')
# log
total_lost += lost_number
print('[{:d}/{:d}] | video: {:12s} | time: {:4.1f}s | speed: {:3.1f}fps | lost_number: {:d} ' \
.format(v_idx + 1, len(dataset), video.name, toc, idx / toc, lost_number))
print('total_lost: {}'.format(total_lost))
def main():
os.environ['CUDA_VISIBLE_DEVICES'] = '{}'.format(args.gpuid)
snap_shot = './checkpoints/model0_e19.pth' #
cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(cur_dir, '../../pysot/testing_dataset', args.dataset) #'LaSOT')#
# create model
model = ManModelBuilder(out_ch=1024, relu=True).cuda()
# load model
model = load_pretrain(model, snap_shot)
torch.set_grad_enabled(False)
# build tracker
tracker = ManTracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=True)
model_name = snap_shot.split('/')[-1].split('.')[0]
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video and args.video != '%d' % v_idx:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [gt_bbox[0], gt_bbox[1],
gt_bbox[0], gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_min_max_bbox(np.array(gt_bbox)) # get_axis_aligned_bbox
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
overlap = vot_overlap(pred_bbox, gt_bbox, (img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(img, [np.array(gt_bbox, np.int).reshape((-1, 1, 2))],
True, (0, 255, 0), 3)
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'.format(
v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video and args.video != '%d' % v_idx:
continue
if 'LaSOT' in args.dataset:
model_path = os.path.join('results', args.dataset, model_name)
result_path = os.path.join(model_path, '{}.txt'.format(video.name))
if os.path.exists(result_path):
print("pass " + video.name)
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
video.load_img()
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - w / 2, cy - h/ 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append((cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
if not np.isnan(gt_bbox).any():
gt_bbox = list(map(int, gt_bbox))
cv2.rectangle(img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]), (0, 255, 0), 3)
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2], pred_bbox[1] + pred_bbox[3]), (0, 0, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, cv2.resize(img, (480, 360)))
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
video.free_img()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write("{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name, video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('results', args.dataset, model_name)#, '{}'.format(video.name))
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path, '{}.txt'.format(video.name))
# result_path = os.path.join(model_path, '{}_001.txt'.format(video.name))
# model_path2 = os.path.join('results', args.dataset, model_name)
# shutil.copyfile(result_path, result_path2)
# if os.path.exists(result_path2):
# print("success copy" + video.name)
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.format(
v_idx + 1, video.name, toc, idx / toc))
def run_tracker(tracker, gt, video_name, restart=True):
frame_count = 0
lost_number = 0
pred_bboxes = []
toc = 0
if restart:
for idx, (frame, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == frame_count:
tracker.init(frame, gt_bbox) # cx,cy,w,h
pred_bboxes.append(1)
elif idx > frame_count:
track_result = tracker.track(frame)
bbox = track_result['bbox'] # cx,cy,w,h
score = track_result['score']
bbox_ = [bbox[0] - bbox[2] / 2, bbox[1] - bbox[3] / 2, bbox[2], bbox[3]] # x,y,w,h
gt_bbox_ = [gt_bbox[0] - gt_bbox[2] / 2, gt_bbox[1] - gt_bbox[3] / 2, gt_bbox[2], gt_bbox[3]]
if vot_overlap(bbox_, gt_bbox_, (frame.shape[1], frame.shape[0])) > 0:
pred_bboxes.append(bbox_)
else:
pred_bboxes.append(2)
frame_count = idx + 5
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if args.vis and idx > frame_count:
show_double_bbox(frame, bbox, score, gt_bbox, idx, lost_number)
toc /= cv2.getTickFrequency()
# log
print('video: {}, time: {:.1f}s, speed: {:.1f}fps, lost_number: {:d} '.format(video_name,
toc, idx / toc,
lost_number))
return pred_bboxes
else:
# toc = 0
# pred_bboxes = []
# scores = []
# track_times = []
# for idx, (img, gt_bbox) in enumerate(video):
# tic = cv2.getTickCount()
# if idx == 0:
# cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
# gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
# tracker.init(img, gt_bbox_)
# pred_bbox = gt_bbox_
# scores.append(None)
# pred_bboxes.append(pred_bbox)
# else:
# outputs = tracker.track(img)
# pred_bbox = outputs['bbox']
# pred_bboxes.append(pred_bbox)
# scores.append(outputs['best_score'])
# toc += cv2.getTickCount() - tic
# track_times.append((cv2.getTickCount() - tic) / cv2.getTickFrequency())
# toc /= cv2.getTickFrequency()
# print('Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.format(
# video_name, toc, idx / toc))
# return pred_bboxes, scores, track_times
pass
def main():
# refine_method = args.refine_method
model_name = 'siamrpn_' + refine_method
model_path = '/'
snapshot_path = os.path.join(
project_path_, 'experiments/%s/model.pth' % args.tracker_name)
config_path = os.path.join(
project_path_, 'experiments/%s/config.yaml' % args.tracker_name)
cfg.merge_from_file(config_path)
dataset_root = dataset_root_
# create model
'''a model is a Neural Network.(a torch.nn.Module)'''
model = ModelBuilder()
# load model
model = load_pretrain(model, snapshot_path).cuda().eval()
# build tracker
'''a tracker is a object, which consists of not only a NN but also some post-processing'''
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
'''##### build a refinement module #####'''
if 'RF' in refine_method:
RF_module = RefineModule(refine_path,
selector_path,
branches=branches,
search_factor=sr,
input_sz=input_sz)
elif refine_method == 'iou_net':
RF_info = Tracker('iou_net', 'iou_net_dimp', None)
RF_params = RF_info.get_parameters()
RF_params.visualization = False
RF_params.debug = False
RF_params.visdom_info = {
'use_visdom': False,
'server': '127.0.0.1',
'port': 8097
}
RF_module = RF_info.tracker_class(RF_params)
elif refine_method == 'mask':
RF_module = siammask()
else:
raise ValueError("refine_method should be 'RF' or 'iou' or 'mask' ")
# model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
H, W, _ = img.shape
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
'''##### initilize refinement module for specific video'''
if 'RF' in refine_method:
RF_module.initialize(
cv2.cvtColor(img, cv2.COLOR_BGR2RGB),
np.array(gt_bbox_))
elif refine_method == 'iou_net':
gt_bbox_np = np.array(gt_bbox_)
gt_bbox_torch = torch.from_numpy(
gt_bbox_np.astype(np.float32))
init_info = {}
init_info['init_bbox'] = gt_bbox_torch
RF_module.initialize(
cv2.cvtColor(img, cv2.COLOR_BGR2RGB), init_info)
elif refine_method == 'mask':
RF_module.initialize(img, np.array(gt_bbox_))
else:
raise ValueError(
"refine_method should be 'RF' or 'RF_mask' or 'iou_net' or 'mask' "
)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
'''##### refine tracking results #####'''
if 'RF' in refine_method or refine_method == 'iou_net':
pred_bbox = RF_module.refine(
cv2.cvtColor(img, cv2.COLOR_BGR2RGB),
np.array(pred_bbox))
x1, y1, w, h = pred_bbox.tolist()
'''add boundary and min size limit'''
x1, y1, x2, y2 = bbox_clip(x1, y1, x1 + w, y1 + h,
(H, W))
w = x2 - x1
h = y2 - y1
pred_bbox = np.array([x1, y1, w, h])
'''pass new state back to base tracker'''
tracker.center_pos = np.array([x1 + w / 2, y1 + h / 2])
tracker.size = np.array([w, h])
elif refine_method == 'mask':
pred_bbox, center_pos, size = RF_module.refine(
img, np.array(pred_bbox), VOT=True)
# boundary and min size limit have been included in "refine"
'''pass new state back to base tracker'''
'''pred_bbox is a list with 8 elements'''
tracker.center_pos = center_pos
tracker.size = size
else:
raise ValueError(
'refine_method should be RF or iou or mask')
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
if refine_method == 'mask':
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join(save_dir, args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
print(
'({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if video.name + '.txt' in os.listdir(model_path):
continue
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
H, W, _ = img.shape
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
'''##### initilize refinement module for specific video'''
if 'RF' in refine_method:
RF_module.initialize(
cv2.cvtColor(img, cv2.COLOR_BGR2RGB),
np.array(gt_bbox_))
elif refine_method == 'iou_net':
gt_bbox_np = np.array(gt_bbox_)
gt_bbox_torch = torch.from_numpy(
gt_bbox_np.astype(np.float32))
init_info = {}
init_info['init_bbox'] = gt_bbox_torch
RF_module.initialize(
cv2.cvtColor(img, cv2.COLOR_BGR2RGB), init_info)
elif refine_method == 'mask':
RF_module.initialize(img, np.array(gt_bbox_))
else:
raise ValueError(
"refine_method should be 'RF' or 'iou' or 'mask' ")
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
'''##### refine tracking results #####'''
if 'RF' in refine_method or refine_method == 'iou_net':
pred_bbox = RF_module.refine(
cv2.cvtColor(img, cv2.COLOR_BGR2RGB),
np.array(pred_bbox))
elif refine_method == 'mask':
pred_bbox = RF_module.refine(img,
np.array(pred_bbox),
VOT=False)
else:
raise ValueError(
"refine_method should be 'RF' or 'iou' or 'mask' ")
x1, y1, w, h = pred_bbox.tolist()
'''add boundary and min size limit'''
x1, y1, x2, y2 = bbox_clip(x1, y1, x1 + w, y1 + h, (H, W))
w = x2 - x1
h = y2 - y1
pred_bbox = np.array([x1, y1, w, h])
tracker.center_pos = np.array([x1 + w / 2, y1 + h / 2])
tracker.size = np.array([w, h])
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(
img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]),
(0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3]), (0, 255, 255),
3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join(save_dir, args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join(save_dir, args.dataset, model_name,
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join(
save_dir, args.dataset,
model_name + '_' + str(selector_path))
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.
format(v_idx + 1, video.name, toc, idx / toc))
def main():
# load config
cfg.merge_from_file(args.config)
cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(cur_dir, '../testing_dataset', args.dataset)
# create model
model = Model2021()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracke
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
overlaps1 = []
vars1 = []
vars0 = []
occl1 = []
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
frame_width = 960#img.shape[1]
frame_height = 540#img.shape[0]
video_loc = os.path.join('../results', model_name, video.name)
out = cv2.VideoWriter(video_loc+'.avi',cv2.VideoWriter_fourcc('M','J','P','G'), 10, (frame_width,frame_height),True)
if video.tags['occlusion']==[] or (np.array(video.tags['occlusion'])==1).sum()==0:
print("\t\tdiscard occlusion")
continue
video.tags['occlusion'] = video.tags['all']
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [gt_bbox[0], gt_bbox[1],
gt_bbox[0], gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx-(w-1)/2, cy-(h-1)/2, w, h]
box1 = gt_bbox_
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
if idx == 0:
print(img.shape)
elif idx > frame_counter:
outputs = tracker.track(img, mode)
pred_bbox = outputs['bbox']
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox, (img.shape[1], img.shape[0]))
#######################################################################################
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx-(w-1)/2, cy-(h-1)/2, w, h]
box2 = gt_bbox_
w1, h1 = box1[2], box1[3]
w2, h2 = box2[2], box2[3]
cx1, cy1 = (img.shape[1]//2, img.shape[0]//2)
cx2, cy2 = (box2[2]/2+box2[0], box2[3]/2+box2[1])
# box1 = box2
# scale variation
s1 = np.sqrt(w1*h1)
s2 = np.sqrt(w2*h2)
sv = max(s1/s2, s2/s1)
# aspect ratio variation
r1, r2 = h1/w1, h2/w2
arv = max(r1/r2, r2/r1)
# fast motion
fm = np.sqrt((cx2-cx1)**2+(cy2-cy1)**2)/np.sqrt(s1*s2)
vars0.append(np.array([sv, arv, fm, outputs['cls2']]))
# occlusion
#########################################################################################
# print(idx, outputs['var'], np.array([sv, arv, fm])) ##################################
overlaps1.append(overlap)
vars1.append(outputs['cls2'])
if idx<=len(video.tags['occlusion']):
occl1.append(video.tags['occlusion'][idx])
else:
occl1.append(np.zeros(idx-len(video.tags['occlusion'])))
if overlap > 0.0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
# print("-------loss---------")
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
for l in range(0,5):
vars1.append(-0.2)
occl1.append(-0.2)
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(img, [np.array(gt_bbox, np.int).reshape((-1, 1, 2))],
True, (255, 0, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(img, [np.array(pred_bbox, np.int).reshape((-1, 1, 2))],
True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0]+bbox[2], bbox[1]+bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
print(idx)
cv2.putText(img, 'occl_gt:'+str(video.tags['occlusion'][idx-1]), (40, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 0, 0), 2)
cv2.putText(img, 'proposed_TL:'+str(lost_number), (40, 160), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, 'occl_pred:'+str(vars1[idx-1]), (40, 120), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
out.write(img)
cv2.imwrite(video_loc+str(idx)+'.png',img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
out.release()
video_path = os.path.join(args.results, args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i) for i in x])+'\n')
print('({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}, mIOU: {:0.4f}'.format(
v_idx+1, video.name, toc, idx / toc, lost_number, np.array(overlaps1).mean()))
# plt.plot(overlaps1)
# plt.plot(np.array(vars0)[:,3])
# plt.plot(np.array(occl1))
# plt.plot(np.array(vars1))
# print(np.correlate(overlaps1,np.array(vars1)[:,2]))
overlaps2.append(np.array(overlaps1).mean())
# occl2.append(np.array(occl1))
# vars2.append(np.array(vars1))
# if args.video != '':
# v_idx=0
# print(100*(confusion_matrix(occl2[v_idx],vars2[v_idx]).ravel()))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
# cv2.destroyAllWindows()
# print("Total Mean IOU is %0.4f"%np.array(overlaps2).mean())
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx-(w-1)/2, cy-(h-1)/2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append((cv2.getTickCount() - tic)/cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0]+gt_bbox[2], gt_bbox[1]+gt_bbox[3]), (0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0]+pred_bbox[2], pred_bbox[1]+pred_bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.rewaitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('../results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write("{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('../results', args.dataset, model_name, video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('../results', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path, '{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.format(
v_idx+1, video.name, toc, idx / toc ))
def main(frame_interval, interpolation_rate):
# load config
cfg.merge_from_file(args.config)
cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(cur_dir, '../testing_dataset', args.dataset)
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 0, 255), 3)
if cfg.MASK.MASK:
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (255, 0, 0), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
window_name = "Result"
cv2.moveWindow(window_name, 100, 100)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# # save results
# video_path = os.path.join('results', args.dataset, model_name,
# 'baseline', video.name)
# if not os.path.isdir(video_path):
# os.makedirs(video_path)
# result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
# with open(result_path, 'w') as f:
# for x in pred_bboxes:
# if isinstance(x, int):
# f.write("{:d}\n".format(x))
# else:
# f.write(','.join([vot_float2str("%.4f", i) for i in x])+'\n')
# print('({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'.format(
# v_idx+1, video.name, toc, idx / toc, lost_number))
# total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# FPS List
fps_list = []
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
# PARAMETERS
# frame_interval = 2
# interpolation_rate = 0.005
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w,
h] # (left-top width height)
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox'] # (left-top width height)
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
######################################
# Adaptive Template(exemplar update) #
######################################
if idx % frame_interval == 0:
tracker.update_z(img,
pred_bbox,
interpolation_rate=interpolation_rate)
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(
img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]),
(0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3]), (255, 0, 0),
3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
window_name = "Result"
cv2.moveWindow(window_name, 20, 20)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
# OTB-100 HERE!!!!!!!!!!!!!!
else:
result_folder_name = "results_{0:d}frame_exemplar_update_rate_{1:s}".format(
frame_interval, str(interpolation_rate))
model_path = os.path.join(result_save_base_path,
result_folder_name, args.dataset,
model_name)
# model_path = os.path.join(result_save_base_path, 'results', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.
format(v_idx + 1, video.name, toc, idx / toc))
# FPS Result
fps = idx / toc
fps_list.append(fps)
# Make FPS Result Path
fps_array = np.asarray(fps_list).reshape(-1, 1)
fps_file_name = "model_fps__[{:3.1f}].txt".format(
np.average(fps_array))
model_fps_file = os.path.join(os.path.dirname(model_path), "../",
fps_file_name)
np.savetxt(model_fps_file, fps_array)
def main():
# load config
cfg.merge_from_file(args.config)
#!!! input your dataset path
dataset_root = os.path.join(your_dataset_path, args.dataset)
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.snapshot.split('/')[-1].split('.')[0]
model_name = model_name + '_pk-{:.3f}'.format(
cfg.TRACK.PENALTY_K) + '_wi-{:.3f}'.format(
cfg.TRACK.WINDOW_INFLUENCE) + '_lr-{:.3f}'.format(cfg.TRACK.LR)
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
print(
'({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('results', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.
format(v_idx + 1, video.name, toc, idx / toc))
def main():
'''change save_path to yours'''
save_path = '/home/masterbin-iiau/Desktop/AdvTrack-project/supplementary/%s' % args.video
if not os.path.exists(save_path):
os.mkdir(save_path)
# load config
cfg.merge_from_file(args.config)
dataset_root = os.path.join(dataset_root_, args.dataset)
# create model
'''a model is a Neural Network.(a torch.nn.Module)'''
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
'''a tracker is a object, which consists of not only a NN but also some post-processing'''
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
# model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
'''GAN'''
outputs = tracker.track_supp(img, GAN, save_path, idx)
pred_bbox = outputs['bbox']
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track_supp(img, GAN, save_path, idx)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(
img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]),
(0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3]), (0, 255, 255),
3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
def main():
net = models.__dict__[args.arch](anchors_nums=args.anchor_nums,
cls_type=args.cls_type)
net = load_pretrain(net, args.resume)
net.eval()
net = net.cuda()
# prepare tracker
info = edict()
info.arch = args.arch
info.cls_type = args.cls_type
info.dataset = args.dataset
info.epoch_test = args.epoch_test
tracker = SiamRPN(info)
dataset_root = os.path.join("/ssd", args.dataset)
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.resume.split('/')[-1].split('.')[0]
total_lost = 0
"""
eao will lower than origin version(0.393->0.390) due to the
number of digits after the decimal point
"""
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
# if len(gt_bbox) == 4:
# gt_bbox = [gt_bbox[0], gt_bbox[1],
# gt_bbox[0], gt_bbox[1]+gt_bbox[3]-1,
# gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]+gt_bbox[3]-1,
# gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
#gt_bbox_ = [cx-(w-1)/2, cy-(h-1)/2, w, h]
target_pos = np.array([cx, cy])
target_sz = np.array([w, h])
state = tracker.init(img, target_pos, target_sz,
net) # init tracker
state["arch"] = args.arch
#tracker.init(img, gt_bbox_)
#pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
state = tracker.track(state, img) # track
location = cxy_wh_2_rect(state['target_pos'],
state['target_sz'])
#outputs = tracker.track(img)
pred_bbox = location
#overlap=poly_iou(gt_bbox,location)
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
print(
'({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
target_pos = np.array([cx, cy])
target_sz = np.array([w, h])
state = tracker.init(img, target_pos, target_sz,
net) # init tracker
state["arch"] = args.arch
#tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
state = tracker.track(state, img) # track
location = cxy_wh_2_rect(state['target_pos'],
state['target_sz'])
pred_bbox = location
#outputs = tracker.track(img)
#pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(state['score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(
img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]),
(0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3]), (0, 255, 255),
3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('result', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.
format(v_idx + 1, video.name, toc, idx / toc))
def main():
# load config
cfg.merge_from_file(args.config)
cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(args.dataset_dir, args.dataset)
epsilon = args.epsilon
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().train()
# build tracker
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False,
config=cfg)
#
# vid.name = {'ants1','ants3',....}
# img, bbox, cls, delta, delta_weight
# vid[0][0],vid[0][1],vid[0][2],vid[0][3],vid[0][4]
model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
# set writing video parameters
height, width, channels = video[0][0].shape
out = cv2.VideoWriter(
os.path.join(args.savedir, video.name + '.avi'),
cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'), 15,
(width, height))
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
data = {'template': None, 'search': None}
for idx, (img, gt_bbox, cls, delta_cls, delta_w, _bbox, cls_s, delta_cls_s, delta_w_s, _bbox_s) \
in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
nimg, sz, box, _ = tracker.crop(img,
bbox=gt_bbox_,
im_name='exemplar')
data['template'] = torch.autograd.Variable(
nimg, requires_grad=True).cuda()
elif idx > frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
nimg, sz, box, pad = tracker.crop(img,
bbox=gt_bbox_,
is_template=False,
im_name='search' +
str(idx))
[bT, bB, bL, bR] = box
sz = int(sz)
data['search'] = torch.autograd.Variable(
nimg, requires_grad=True).cuda()
data['label_cls'] = torch.Tensor(cls_s).type(
torch.LongTensor).cuda()
data['label_loc'] = torch.Tensor(delta_cls_s).type(
torch.FloatTensor).cuda()
data['label_loc_weight'] = torch.Tensor(delta_w_s).cuda()
outputs = model(data)
cls_loss = outputs['cls_loss']
loc_loss = outputs['loc_loss']
total_loss = outputs['total_loss']
total_loss.backward()
data_grad = data['search'].grad
# torch.Tensor(img.transpose([2, 0, 1])).unsqueeze(dim=0)
perturb_data = fgsm_attack(data['search'], epsilon,
data_grad)
# cv2.imwrite(os.path.join(args.savedir, 'original_' + str(idx) + '.jpg'), img)
# _img = perturb_data.data.cpu().numpy().squeeze().transpose([1, 2, 0])
# cv2.imwrite(os.path.join(args.savedir, 'perturb_' + str(idx) + '.jpg'), _img)
if not np.array_equal(cfg.TRACK.INSTANCE_SIZE, sz):
perturb_data = F.interpolate(perturb_data, size=sz)
_img = perturb_data.data.cpu().numpy().squeeze().transpose(
[1, 2, 0])
# cv2.imwrite(os.path.join(args.savedir, 'crop_full_' + str(idx) + '.jpg'), _img)
nh, nw, _ = _img.shape
img[bT:bB + 1, bL:bR + 1, :] = _img[pad[0]:nh - pad[1],
pad[2]:nw - pad[3], :]
# cv2.imwrite(os.path.join(args.savedir, 'perturb_full_' + str(idx) + '.jpg'), img)
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
print('*************** lost ***************')
import pdb
pdb.set_trace()
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
print(idx, torch.sum(data_grad, (2, 3)))
print(
idx,
torch.sum(torch.abs(torch.sum(data_grad, (2, 3))),
(0, 1)))
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
# save tracking image
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX,
1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
# cv2.imwrite(os.path.join(args.savedir, 'track_' + str(idx) + '.jpg'), img)
out.write(img)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
print(
'({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(
img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]),
(0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3]), (0, 255, 255),
3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('results', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.
format(v_idx + 1, video.name, toc, idx / toc))
def main():
# load config
# save_siamese_rpn()
cfg.merge_from_file(args.config)
cur_dir = os.path.dirname(os.path.realpath(__file__))
# dataset_root = os.path.join(cur_dir, '../testing_dataset', args.dataset)
dataset_root = datasets_root + args.dataset
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# save_backbone(model)
# build tracker
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
#multi-pass tracking,跟踪丢失后重新初始化的测试方法
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
# pred_bboxes包含两种类型的数据，类型1：整型数据，有1,2，0,三个值，分别表示跟踪开始，跟踪结束（丢失），跟踪丢失之后，间隔帧的占位符
# 类型2：浮点类型的bbox,也就是跟踪结果
pred_bboxes = []
gru_seq_len = tracker.model.grus.seq_in_len
video_len = len(video)
for idx, (img, gt_bbox) in enumerate(video):
if len(
gt_bbox
) == 4: #如果gt是【x，y,w,h】的方式，转化为8个坐标信息（x1,y1,x2,y2,x3,y3,x4,y4）
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
#跟踪初始化
if idx == frame_counter: # 跟踪第一帧初始化
idxs = list(
map(lambda x, y: x + y, [idx] * gru_seq_len,
list(range(
0,
gru_seq_len)))) # 取出idx后面的gru_seq_len个序列的索引号
idxs = list(map(lambda x: min(x, video_len - 1),
idxs)) # 避免索引号越界
tracker.template_idx = 0 #模板初始化的第一帧
for k in idxs:
init_img, init_gt_bbox = video[k] #连续gru_seq_len帧初始化
#init_img, init_gt_bbox =video[idxs[0]] #只用一帧作为初始化参数
cx, cy, w, h = get_axis_aligned_bbox(
np.array(init_gt_bbox)
) #将倾斜框4个点坐标，转化为bbox,x,y为中心点形式(cx,cy,w,h)
init_gt_bbox = [
cx - (w - 1) / 2, cy - (h - 1) / 2, w, h
] #x,y,中心点形式，转化为左上角形式
tracker.init_gru(init_img, init_gt_bbox, k)
if k == 0:
pred_bbox = init_gt_bbox
pred_bboxes.append(1)
#持续的后续跟踪
elif idx > frame_counter:
outputs = tracker.track(img) #对于下面的帧
pred_bbox = outputs['bbox']
#只有输出概率很高的时候才更新模板
if outputs['best_score'] > 0.95:
tracker.init_gru(img, pred_bbox, idx)
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
#查看初始化后的第一帧检测iou和score之间的关系
# if tracker.template_idx==4:
# print("{:3.2f}\t{:3.2f}".format(overlap,outputs['best_score']))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
#绘制输出框，gt和mask都按照多边形来绘制，跟踪的bbox按照矩形来绘制
if args.vis and idx > frame_counter:
#绘制多边形的gt
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
#绘制siamesemask输出的多边形
if cfg.MASK.MASK:
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 255), 3)
#绘制输出矩形框
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
#添加图像标注，帧号和丢失次数
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
#结果路径的构成： ./results/VOT2018/model/baseline/ants1/ants1_001.txt
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
#pred_bboxes包含两种类型的数据，类型1：整型数据，有1,2，0,三个值，分别表示跟踪开始，跟踪结束（丢失），跟踪丢失之后，间隔帧的占位符
# 类型2：浮点类型的bbox,也就是跟踪结果
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int): #整数代表开始，或者有丢失
f.write("{:d}\n".format(x))
else: #浮点数才是bbox
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
print(
'({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
#oetracking,跟踪丢失后不再重新初始化的测试方法
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(
img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]),
(0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3]), (0, 255, 255),
3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('results', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.
format(v_idx + 1, video.name, toc, idx / toc))
def main():
# load config
cfg.merge_from_file(args.config)
# -------------------------------------hp_search---------------------------------------#
params = [0.0, 0.0, 0.0]
# Interpolation learning rate
params[0] = cfg.TRACK.LR
# Scale penalty
params[1] = cfg.TRACK.PENALTY_K
# Window influence
params[2] = cfg.TRACK.WINDOW_INFLUENCE
params_name = args.snapshot.split(
'/')[-1] + ' ' + args.dataset + ' lr-' + str(
params[0]) + ' pk-' + '_' + str(params[1]) + ' win-' + '_' + str(
params[2])
# -------------------------------------hp_search---------------------------------------#
# cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join('./datasets', args.dataset)
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker siamos
tracker = SiamCARTracker(model, cfg.TRACK)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
total_lost = 0
avg_speed = 0 # linlin
for v_idx, video in tqdm(enumerate(dataset)):
#for v_idx, video in tqdm(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w,
h] #[topx,topy,w,h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join(args.save_path, args.dataset,
args.tracker_name, 'baseline',
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
# print('({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'.format(
# v_idx+1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
avg_speed += idx / toc
print('Speed: {:3.1f}fps'.format(avg_speed / 60))
print(params_name)
#print(" stage:{:d} model:{:s} epoch:{:s} update_lr:{:f}".format(args.update_stage,args.update_path, args.update_path.split('/')[-1],update_lr[args.update_lr]))
else:
# OPE tracking
for v_idx, video in tqdm(enumerate(dataset)):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w,
h] #[topx,topy,w,h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
#scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(
img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]),
(0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3]), (0, 255, 255),
3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join(args.save_path, args.dataset,
args.tracker_name, 'longterm',
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join(args.save_path, args.dataset,
args.tracker_name, video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join(args.save_path, args.dataset,
args.tracker_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
# print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.format(
# v_idx+1, video.name, toc, idx / toc))
print(params_name)
# os.chdir(model_path)
# save_file = '../%s' % dataset
# shutil.make_archive(save_file, 'zip')
#print('Records saved at', save_file + '.zip')
evaluate(args)
def main(args, tracker):
# create dataset
if not args.dataset_path:
args.dataset_path = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(args.dataset_path, 'dataset', args.dataset)
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False,
single_video=args.video)
model_name = args.model_name
if args.debug_vis:
args.vis = True
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019', 'VOT2020']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
video_total_lost = 0
for cnt in range(args.repetition):
frame_counter = 0
lost_number = 0
toc = 0
init_toc = 0
valid_frames = 0
pred_bboxes = []
template_image = None
search_image = None
raw_heatmap = None
post_heatmap = None
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [gt_bbox[0], gt_bbox[1],
gt_bbox[0], gt_bbox[1]+gt_bbox[3],
gt_bbox[0]+gt_bbox[2], gt_bbox[1]+gt_bbox[3],
gt_bbox[0]+gt_bbox[2], gt_bbox[1]]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - w/2, cy - h/2, w, h]
tracker.init(img, gt_bbox_)
init_toc += cv2.getTickCount() - tic
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bbox = [pred_bbox[0], pred_bbox[1],
pred_bbox[2] - pred_bbox[0],
pred_bbox[3] - pred_bbox[1]]
valid_frames += 1
toc += cv2.getTickCount() - tic
overlap = vot_overlap(pred_bbox, gt_bbox, (img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
if args.vis and args.debug_vis:
cv2.polylines(img, [np.array(gt_bbox, np.int).reshape((-1, 2))], True, (0, 255, 0), 3)
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0]+bbox[2], bbox[1]+bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, 'lost', (40, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
for key, value in outputs.items():
if isinstance(value, np.ndarray):
if len(value.shape) == 3 or len(value.shape) == 2:
cv2.imshow(key, value)
k = cv2.waitKey(0)
if k == 27: # wait for ESC key to exit
sys.exit()
else:
pred_bboxes.append(0)
if idx == 0:
if args.vis:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(img, [np.array(gt_bbox, np.int).reshape((-1, 2))], True, (0, 255, 0), 3)
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0]+bbox[2], bbox[1]+bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
if args.debug_vis:
for key, value in outputs.items():
if isinstance(value, np.ndarray):
if len(value.shape) == 3 or len(value.shape) == 2:
cv2.imshow(key, value)
k = cv2.waitKey(0)
if k == 27: # wait for ESC key to exit
break
else:
k = cv2.waitKey(1)
if k == 27: # wait for ESC key to exit
break
sys.stderr.write("inference on {}: {} / {}\r".format(video.name, idx+1, len(video)))
toc /= cv2.getTickFrequency()
init_toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join(args.result_path, args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_{:03d}.txt'.format(video.name, cnt+1))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i) for i in x])+'\n')
log = '({:3d}) Video ({:2d}): {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'.format(
v_idx+1, cnt+1, video.name, init_toc + toc, valid_frames / toc, lost_number)
print(log)
with open(os.path.join(args.result_path, args.dataset, model_name, 'log.txt'), 'a') as f:
f.write(log + '\n')
video_total_lost += lost_number
total_lost += video_total_lost
if args.repetition > 1:
log = '({:3d}) Video: {:12s} Avg Lost: {:.3f}'.format(v_idx+1, video.name, video_total_lost/args.repetition)
print(log)
with open(os.path.join(args.result_path, args.dataset, model_name, 'log.txt'), 'a') as f:
f.write(log + '\n')
log = "{:s} total (avg) lost: {:.3f}".format(model_name, total_lost/args.repetition)
print(log)
with open(os.path.join(args.result_path, args.dataset, model_name, 'log.txt'), 'a') as f:
f.write(log + '\n')
else:
# OPE tracking
find_best = True
if not dataset.has_ground_truth:
find_best = False
# if repeat 3 times for GOT-10k, use the official benchmark mode (no find best)
if args.dataset == 'GOT-10k':
if args.repetition == 3:
find_best = False
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
best_pred_bboxes = []
min_lost_number = 1e6
for cnt in range(args.repetition):
toc = 0
init_toc = 0
pred_bboxes = []
track_times = []
template_image = None
search_image = None
raw_heatmap = None
post_heatmap = None
lost_number = 0
if find_best and min_lost_number < args.min_lost_rate_for_repeat * len(video):
print("Abolish reset of trails ({}~) becuase the min lost number is small enough: {} / {}".format(cnt+1 , min_lost_number, args.min_lost_rate_for_repeat * len(video)))
break
save_image_offset = 0
if args.save_image_num_per_video > 1:
save_image_offset = len(video) // (args.save_image_num_per_video - 1)
if args.save_image_num_per_video == 0:
save_image_offset = 1
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
outputs = tracker.init(img, gt_bbox)
init_toc += cv2.getTickCount() - tic
pred_bbox = gt_bbox
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
toc += cv2.getTickCount() - tic
pred_bbox_ = outputs['bbox']
pred_bbox = [pred_bbox_[0], pred_bbox_[1],
pred_bbox_[2] - pred_bbox_[0],
pred_bbox_[3] - pred_bbox_[1]]
pred_bboxes.append(pred_bbox)
track_times.append((cv2.getTickCount() - tic)/cv2.getTickFrequency())
gt_bbox_int = list(map(lambda x: int(x) if not np.isnan(x) else 0, gt_bbox))
pred_bbox_int = list(map(int, pred_bbox))
cv2.rectangle(img, (gt_bbox_int[0], gt_bbox_int[1]),
(gt_bbox_int[0]+gt_bbox_int[2], gt_bbox_int[1]+gt_bbox_int[3]), (0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox_int[0], pred_bbox_int[1]),
(pred_bbox_int[0]+pred_bbox_int[2], pred_bbox_int[1]+pred_bbox_int[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
if save_image_offset > 0:
image_path = os.path.join(args.result_path, args.dataset, model_name, 'images', video.name)
if not os.path.isdir(image_path):
os.makedirs(image_path)
if idx % save_image_offset == 0:
imagename = os.path.join(image_path, 'image{:03d}.jpg'.format(idx))
cv2.imwrite(imagename,img)
if idx == 0:
if args.vis:
cv2.destroyAllWindows()
if args.debug_vis and isinstance(outputs, dict):
for key, value in outputs.items():
if isinstance(value, np.ndarray):
if len(value.shape) == 3 or len(value.shape) == 2:
cv2.imshow(key, value)
else:
if not gt_bbox == [0,0,0,0] and not np.isnan(np.array(gt_bbox)).any():
if pred_bbox[0] + pred_bbox[2] < gt_bbox[0] or pred_bbox[0] > gt_bbox[0] + gt_bbox[2] or pred_bbox[1] + pred_bbox[3] < gt_bbox[1] or pred_bbox[1] > gt_bbox[1] + gt_bbox[3]:
lost_number += 1
if find_best and lost_number > min_lost_number:
break
if args.vis or args.debug_vis:
cv2.imshow(video.name, img)
if args.debug_vis:
for key, value in outputs.items():
if isinstance(value, np.ndarray):
if len(value.shape) == 3 or len(value.shape) == 2:
cv2.imshow(key, value)
k = cv2.waitKey(0)
if k == 27: # wait for ESC key to exit
min_lost_number = 1e6 # this allows to try args.repetition times for debug
lost_number = 1e6 # this allows to try args.repetition times for debug
break
else:
k = cv2.waitKey(1)
if k == 27: # wait for ESC key to exit
min_lost_number = 1e6 # this allows to try args.repetition times for debug
lost_number = 1e6 # this allows to try args.repetition times for debug
break
sys.stderr.write("inference on {}: {} / {}\r".format(video.name, idx+1, len(video)))
if find_best and lost_number > min_lost_number:
print('Stop No.{} trial becuase the lost number already exceed the min lost number: {} > {} '.format(cnt+1, lost_number, min_lost_number))
continue
if lost_number == 1e6:
continue
if lost_number < min_lost_number:
min_lost_number = lost_number
toc /= cv2.getTickFrequency()
init_toc /= cv2.getTickFrequency()
# save results
if 'GOT-10k' == args.dataset:
video_path = os.path.join(args.result_path, args.dataset, model_name, video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
id = cnt + 1
if find_best:
id = 1
result_path = os.path.join(video_path, '{}_{:03d}.txt'.format(video.name, id))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([vot_float2str("%.4f", i) for i in x ])+'\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join(args.result_path, args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path, '{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([vot_float2str("%.4f", i) for i in x])+'\n')
log = '({:3d}) Video: {:12s} Trail: {:2d} Time: {:5.1f}s Speed: {:3.1f}fps Lost: {:d}/{:d}'.format(
v_idx+1, video.name, cnt+1, init_toc + toc, idx / toc, lost_number, len(video))
print(log)
with open(os.path.join(args.result_path, args.dataset, model_name, 'log.txt'), 'a') as f:
f.write(log + '\n')
def run_tracker(tracker, img, gt, video_name, video, restart=True):
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
if restart: # VOT2016 and VOT 2018
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1, gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1, gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append([1])
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append([2])
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append([0])
toc += cv2.getTickCount() - tic
toc /= cv2.getTickFrequency()
# print('Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'.format(video_name, toc, idx / toc, lost_number))
return pred_bboxes
else:
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
toc /= cv2.getTickFrequency()
# print('Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.format(video_name, toc, idx / toc))
return pred_bboxes, scores, track_times
def main():
# load config
model_name = 'RT_MDNet_refine'
MASK = False
dataset_root = os.path.join(dataset_root_, args.dataset)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
'''##### build a scale-estimator #####'''
SE_module = Scale_Estimator_bcm(refine_checkpoint_dir_)
# model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [
gt_bbox[0], gt_bbox[1], gt_bbox[0],
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1,
gt_bbox[1] + gt_bbox[3] - 1,
gt_bbox[0] + gt_bbox[2] - 1, gt_bbox[1]
]
tic = cv2.getTickCount()
if idx == frame_counter:
H, W, _ = img.shape
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.init(img, gt_bbox_)
'''##### initilize scale-estimator for specific video'''
# SE_module.initialize(cv2.cvtColor(img,cv2.COLOR_BGR2RGB),
# np.array(gt_bbox_))
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
'''##### refine tracking results #####'''
# output_dict = SE_module.refine_all(cv2.cvtColor(img,cv2.COLOR_BGR2RGB),
# np.array(pred_bbox))
# pred_bbox = 0.5 * (output_dict['bbox'] + output_dict['corner'])
lr = outputs['lr']
pred_bbox = tracker.smooth_bbox(pred_bbox, lr, H, W)
x1, y1, w, h = pred_bbox
tracker.center_pos = np.array([x1 + w / 2, y1 + h / 2])
tracker.size = np.array([w, h])
# if cfg.MASK.MASK:
# pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox,
(img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(
img, [np.array(gt_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 0), 3)
if MASK:
cv2.polylines(
img,
[np.array(pred_bbox, np.int).reshape(
(-1, 1, 2))], True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
print(
'({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'
.format(v_idx + 1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if v_idx < 60:
continue
tracker = RT_MDNet()
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
img_RGB = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # RGB format
tic = cv2.getTickCount()
if idx == 0:
H, W, _ = img.shape
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx - (w - 1) / 2, cy - (h - 1) / 2, w, h]
tracker.initialize_seq(img_RGB, np.array(gt_bbox_))
'''##### initilize scale-estimator for specific video'''
SE_module.initialize(img_RGB, np.array(gt_bbox_))
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
ori_bbox = tracker.track(img_RGB)
'''##### refine tracking results #####'''
output_dict = SE_module.refine_all(img_RGB,
np.array(ori_bbox))
pred_bbox = 0.5 * (output_dict['bbox'] +
output_dict['corner'])
pred_bbox = bbox_clip(pred_bbox, (H, W))
tracker.target_bbox = pred_bbox.copy()
pred_bboxes.append(pred_bbox)
# scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append(
(cv2.getTickCount() - tic) / cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
ori_bbox = list(map(int, ori_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(
img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0] + gt_bbox[2], gt_bbox[1] + gt_bbox[3]),
(0, 0, 255), 3)
cv2.rectangle(
img, (ori_bbox[0], ori_bbox[1]),
(ori_bbox[0] + ori_bbox[2], ori_bbox[1] + ori_bbox[3]),
(255, 0, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0] + pred_bbox[2],
pred_bbox[1] + pred_bbox[3]), (0, 255, 0),
3)
cv2.putText(img, str(idx), (40, 40),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(
video_path, '{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write(
"{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('results', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path,
'{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x]) + '\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.
format(v_idx + 1, video.name, toc, idx / toc))
def main():
# load config
cfg.merge_from_file(args.config)
cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(cur_dir, '../testing_dataset', args.dataset)
# create model
model = ModelBuilder(cfg)
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.snapshot.split('/')[-1].split('.')[0]
total_lost = 0
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
# restart tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [gt_bbox[0], gt_bbox[1],
gt_bbox[0], gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx-(w-1)/2, cy-(h-1)/2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox, (img.shape[1], img.shape[0]))
if overlap > 0:
# not lost
pred_bboxes.append(pred_bbox)
else:
# lost object
pred_bboxes.append(2)
frame_counter = idx + 5 # skip 5 frames
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(img, [np.array(gt_bbox, np.int).reshape((-1, 1, 2))],
True, (0, 255, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(img, [np.array(pred_bbox, np.int).reshape((-1, 1, 2))],
True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0]+bbox[2], bbox[1]+bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join('results', args.dataset, model_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i) for i in x])+'\n')
print('({:3d}) Video: {:12s} Time: {:4.1f}s Speed: {:3.1f}fps Lost: {:d}'.format(
v_idx+1, video.name, toc, idx / toc, lost_number))
total_lost += lost_number
print("{:s} total lost: {:d}".format(model_name, total_lost))
else:
# OPE tracking
for v_idx, video in enumerate(dataset):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx-(w-1)/2, cy-(h-1)/2, w, h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append((cv2.getTickCount() - tic)/cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0]+gt_bbox[2], gt_bbox[1]+gt_bbox[3]), (0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0]+pred_bbox[2], pred_bbox[1]+pred_bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write("{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join('results', args.dataset, model_name, video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join('results', args.dataset, model_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path, '{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
print('({:3d}) Video: {:12s} Time: {:5.1f}s Speed: {:3.1f}fps'.format(
v_idx+1, video.name, toc, idx / toc))
def main():
cfg.merge_from_file(args.config)
dataset_root = os.path.join('./datasets', args.dataset)
params = [0.0,0.0,0.0]
params[0] =cfg.TRACK.LR
params[1]=cfg.TRACK.PENALTY_K
params[2] =cfg.TRACK.WINDOW_INFLUENCE
params_name = args.snapshot.split('/')[-1] + ' '+ args.dataset + ' lr-' + str(params[0]) + ' pk-' + '_' + str(params[1]) + ' win-' + '_' + str(params[2])
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
if args.dataset in ['VOT2016', 'VOT2018', 'VOT2019']:
total_lost=0
avg_speed =0
for v_idx, video in tqdm(enumerate(dataset)):
if args.video != '':
if video.name != args.video:
continue
frame_counter = 0
lost_number = 0
toc = 0
pred_bboxes = []
for idx, (img, gt_bbox) in enumerate(video):
if len(gt_bbox) == 4:
gt_bbox = [gt_bbox[0], gt_bbox[1],
gt_bbox[0], gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]+gt_bbox[3]-1,
gt_bbox[0]+gt_bbox[2]-1, gt_bbox[1]]
tic = cv2.getTickCount()
if idx == frame_counter:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx-(w-1)/2, cy-(h-1)/2, w, h] #[topx,topy,w,h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
pred_bboxes.append(1)
elif idx > frame_counter:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
if cfg.MASK.MASK:
pred_bbox = outputs['polygon']
overlap = vot_overlap(pred_bbox, gt_bbox, (img.shape[1], img.shape[0]))
if overlap > 0:
pred_bboxes.append(pred_bbox)
else:
pred_bboxes.append(2)
frame_counter = idx + 5
lost_number += 1
else:
pred_bboxes.append(0)
toc += cv2.getTickCount() - tic
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > frame_counter:
cv2.polylines(img, [np.array(gt_bbox, np.int).reshape((-1, 1, 2))],
True, (0, 255, 0), 3)
if cfg.MASK.MASK:
cv2.polylines(img, [np.array(pred_bbox, np.int).reshape((-1, 1, 2))],
True, (0, 255, 255), 3)
else:
bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (bbox[0], bbox[1]),
(bbox[0]+bbox[2], bbox[1]+bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.putText(img, str(lost_number), (40, 80), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
video_path = os.path.join(args.save_path, args.dataset, args.tracker_name,
'baseline', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i) for i in x])+'\n')
total_lost += lost_number
avg_speed += idx / toc
print('Speed: {:3.1f}fps'.format(avg_speed/60))
print(params_name)
else:
# OPE tracking
for v_idx, video in tqdm(enumerate(dataset)):
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
for idx, (img, gt_bbox) in enumerate(video):
tic = cv2.getTickCount()
if idx == 0:
cx, cy, w, h = get_axis_aligned_bbox(np.array(gt_bbox))
gt_bbox_ = [cx-(w-1)/2, cy-(h-1)/2, w, h] #[topx,topy,w,h]
tracker.init(img, gt_bbox_)
pred_bbox = gt_bbox_
scores.append(None)
if 'VOT2018-LT' == args.dataset:
pred_bboxes.append([1])
else:
pred_bboxes.append(pred_bbox)
else:
outputs = tracker.track(img)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
#scores.append(outputs['best_score'])
toc += cv2.getTickCount() - tic
track_times.append((cv2.getTickCount() - tic)/cv2.getTickFrequency())
if idx == 0:
cv2.destroyAllWindows()
if args.vis and idx > 0:
gt_bbox = list(map(int, gt_bbox))
pred_bbox = list(map(int, pred_bbox))
cv2.rectangle(img, (gt_bbox[0], gt_bbox[1]),
(gt_bbox[0]+gt_bbox[2], gt_bbox[1]+gt_bbox[3]), (0, 255, 0), 3)
cv2.rectangle(img, (pred_bbox[0], pred_bbox[1]),
(pred_bbox[0]+pred_bbox[2], pred_bbox[1]+pred_bbox[3]), (0, 255, 255), 3)
cv2.putText(img, str(idx), (40, 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(video.name, img)
cv2.waitKey(1)
toc /= cv2.getTickFrequency()
# save results
if 'VOT2018-LT' == args.dataset:
video_path = os.path.join(args.save_path, args.dataset, args.tracker_name,
'longterm', video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path,
'{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_001_confidence.value'.format(video.name))
with open(result_path, 'w') as f:
for x in scores:
f.write('\n') if x is None else f.write("{:.6f}\n".format(x))
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
elif 'GOT-10k' == args.dataset:
video_path = os.path.join(args.save_path, args.dataset, args.tracker_name, video.name)
if not os.path.isdir(video_path):
os.makedirs(video_path)
result_path = os.path.join(video_path, '{}_001.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
result_path = os.path.join(video_path,
'{}_time.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in track_times:
f.write("{:.6f}\n".format(x))
else:
model_path = os.path.join(args.save_path, args.dataset, args.tracker_name)
if not os.path.isdir(model_path):
os.makedirs(model_path)
result_path = os.path.join(model_path, '{}.txt'.format(video.name))
with open(result_path, 'w') as f:
for x in pred_bboxes:
f.write(','.join([str(i) for i in x])+'\n')
eval(args)
