def main():
rank, world_size = dist_init()
# rank = 0
logger.info("init done")
# load cfg
cfg.merge_from_file(args.cfg)
if rank == 0:
if not os.path.exists(cfg.TRAIN.LOG_DIR):
os.makedirs(cfg.TRAIN.LOG_DIR)
init_log('global', logging.INFO)
if cfg.TRAIN.LOG_DIR:
add_file_handler('global',
os.path.join(cfg.TRAIN.LOG_DIR, 'logs.txt'),
logging.INFO)
logger.info("Version Information: \n{}\n".format(commit()))
logger.info("config \n{}".format(json.dumps(cfg, indent=4)))
# create model
model = ModelBuilder().train()
dist_model = nn.DataParallel(model).cuda()
# load pretrained backbone weights
if cfg.BACKBONE.PRETRAINED:
cur_path = os.path.dirname(os.path.realpath(__file__))
backbone_path = os.path.join(cur_path, '../', cfg.BACKBONE.PRETRAINED)
load_pretrain(model.backbone, backbone_path)
# create tensorboard writer
if rank == 0 and cfg.TRAIN.LOG_DIR:
tb_writer = SummaryWriter(cfg.TRAIN.LOG_DIR)
else:
tb_writer = None
# build dataset loader
train_loader = build_data_loader()
# build optimizer and lr_scheduler
optimizer, lr_scheduler = build_opt_lr(dist_model.module,
cfg.TRAIN.START_EPOCH)
# resume training
if cfg.TRAIN.RESUME:
logger.info("resume from {}".format(cfg.TRAIN.RESUME))
assert os.path.isfile(cfg.TRAIN.RESUME), \
'{} is not a valid file.'.format(cfg.TRAIN.RESUME)
model, optimizer, cfg.TRAIN.START_EPOCH = \
restore_from(model, optimizer, cfg.TRAIN.RESUME)
# load pretrain
elif cfg.TRAIN.PRETRAINED:
load_pretrain(model, cfg.TRAIN.PRETRAINED)
dist_model = nn.DataParallel(model)
logger.info(lr_scheduler)
logger.info("model prepare done")
# start training
train(train_loader, dist_model, optimizer, lr_scheduler, tb_writer)
def save_siamese_rpn():
# load config
rpn_path = root_dir + 'experiments/siamrpn_alex_dwxcorr_16gpu/pre_train/checkpoint_e45.pth'
gru_rpn = root_dir + 'experiments/siamrpn_alex_dwxcorr_16gpu/config.yaml'
cfg.merge_from_file(gru_rpn)
# create model
model_rpn = ModelBuilder()
model_rpn = load_pretrain(model_rpn, rpn_path).cuda().eval()
gru_path = root_dir + 'experiments/siamrpn_alex_dwxcorr_16gpu/gru_snapshot/gru_10.pth'
gru_cfg = root_dir + 'experiments/siamrpn_alex_dwxcorr_16gpu/config_gru.yaml'
cfg.merge_from_file(gru_cfg)
# create model
model_gru = ModelBuilder()
model_gru = load_pretrain(model_gru, gru_path).cuda().eval()
for key, item in model_gru.named_parameters():
# print(key.find("grus"))
print(key, item.shape)
for key, item in model_rpn.named_parameters():
# print(key.find("grus"))
print(key, item.shape)
model_gru_dict = model_gru.state_dict()
model_rpn_dict = model_rpn.state_dict()
for key in model_gru_dict:
if key.find("grus") != -1:
print("fix:", key)
else:
print("change:", key)
model_gru_dict[key] = model_rpn_dict[key]
# name_map={}
# model_legacy_dict = model_legacy.state_dict()
# model_alexnet_dict = model_alexnet.state_dict()
# for para1,para2 in zip(model_legacy.named_parameters(),model_alexnet.named_parameters()):
# # print(para1[0],para1[1].shape)
# print(para1[0])
# print(para2[0])
# print(para1[1].shape)
# print(para2[1].shape)
# print("--"*40)
# # print("['{}'--->'{}']".format(para1[0], para2[0]),para1[1].shape, para2[1].shape)
# name_map[para1[0]]=para2[0]
# print(name_map)
#
#
# for key,val in name_map.items():
# model_alexnet_dict[val]=model_legacy_dict[key]
torch.save(model_gru_dict, "siamese_gru10_rpn45.pth")
def main():
rank, world_size = dist_init()
logger.info("init done")
# load cfg
cfg.merge_from_file(args.cfg)
if rank == 0:
if not os.path.exists(cfg.TRAIN.LOG_DIR):
os.makedirs(cfg.TRAIN.LOG_DIR)
init_log('global', logging.INFO)
if cfg.TRAIN.LOG_DIR:
add_file_handler('global',
os.path.join(cfg.TRAIN.LOG_DIR, 'logs.txt'),
logging.INFO)
logger.info("Version Information: \n{}\n".format(commit()))
logger.info("config \n{}".format(json.dumps(cfg, indent=4)))
# create model
model = Template_Enhance().cuda().train()
dist_model = DistModule(model)
# load pretrained SiamRPN++ model
if cfg.TSA.MODELBUILD_PATH:
cur_path = os.path.dirname(os.path.realpath(__file__))
backbone_path = os.path.join(cur_path, '../', cfg.TSA.MODELBUILD_PATH)
load_pretrain(model.model, backbone_path)
# create tensorboard writer
if rank == 0 and cfg.TRAIN.LOG_DIR:
tb_writer = SummaryWriter(cfg.TRAIN.LOG_DIR)
else:
tb_writer = None
# build datasets loader
train_loader, val_loader = build_data_loader()
# build optimizer and lr_scheduler
optimizer, scheduler = build_opt_lr(dist_model.module)
logger.info(scheduler)
logger.info("model prepare done")
# start estimation
# train(train_loader, dist_model, optimizer, scheduler, tb_writer)
estimation = Estimator(train_loader, val_loader, dist_model, optimizer, scheduler, tb_writer)
# estimation.evaluate(5000, 'loss')
estimation.train()
def setup_tracker():
cfg.merge_from_file(cfg_file)
model = ModelBuilder()
model = load_pretrain(model, model_file).cuda().eval()
tracker = build_tracker(model)
warmup(model)
return tracker
def main():
# load config
cfg.merge_from_file(args.config)
cfg.CUDA = torch.cuda.is_available()
device = torch.device('cuda' if cfg.CUDA else 'cpu')
# create model
model = ModelBuilder()
# load model
# model.load_state_dict(torch.load(args.snapshot,
# map_location=lambda storage, loc: storage.cpu()))
# model.eval().to(device)
model = load_pretrain(model, args.snapshot).eval().to(device)
# build tracker
tracker = build_tracker(model)
first_frame = True
if args.video_name:
video_name = args.video_name.split('/')[-1].split('.')[0]
else:
video_name = 'webcam'
cv2.namedWindow(video_name, cv2.WND_PROP_FULLSCREEN)
for frame in get_frames(args.video_name):
if first_frame:
try:
init_rect = cv2.selectROI(video_name, frame, False, False)
except:
exit()
tracker.init(frame, init_rect)
first_frame = False
else:
outputs = tracker.track(frame)
if 'polygon' in outputs:
polygon = np.array(outputs['polygon']).astype(np.int32)
cv2.polylines(frame, [polygon.reshape((-1, 1, 2))], True,
(0, 255, 0), 3)
mask = ((outputs['mask'] > cfg.TRACK.MASK_THERSHOLD) * 255)
mask = mask.astype(np.uint8)
mask = np.stack([mask, mask * 255, mask]).transpose(1, 2, 0)
frame = cv2.addWeighted(frame, 0.77, mask, 0.23, -1)
else:
bbox = list(map(int, outputs['bbox']))
cv2.rectangle(frame, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]),
(0, 255, 0), 3)
cv2.imshow(video_name, frame)
cv2.waitKey(40)
def __init__(self):
super(DROL, self).__init__("DROL")
# load config
cfg.merge_from_file(path_config.DROL_CONFIG)
seed_torch(cfg.TRACK.SEED)
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, path_config.DROL_SNAPSHOT).cuda().eval()
# build tracker
self.tracker = build_tracker(model)
def __init__(self, lr_u=0.2,lr_v=0.2,lambda_u=0.1,lambda_v=10.0,x_padding=0.5, z_ratio=1.2,features='gray', kernel='gaussian'):
super(SFKCF).__init__()
self.x_padding = x_padding
self.lambda_ = 1e-4
self.features = features
self.w2c=None
if self.features=='hog':
self.interp_factor = 0.02
self.sigma = 0.5
self.cell_size=4
self.output_sigma_factor=0.1
elif self.features=='sfres50':
self.interp_factor = 0.02
self.sigma = 0.5
self.cell_size=8.0
self.output_sigma_factor=0.1
model = ModelBuilder()
model = load_pretrain(model, cfg.BACKBONE.PRETRAINED).backbone
self.model = model.cuda().eval()
elif self.features=='gray' or self.features=='color':
self.interp_factor=0.075
self.sigma=0.2
self.cell_size=1
self.output_sigma_factor=0.1
elif self.features=='cn':
self.interp_factor=0.075
self.sigma=0.2
self.cell_size=1
self.output_sigma_factor=1./16
self.padding=1
else:
raise NotImplementedError
self.kernel=kernel
self.U = None
self.V = None
self.lr_u = lr_u
self.lr_v = lr_v
self.lambda_v = lambda_v
self.lambda_u = lambda_u
self.z_padding = z_ratio*x_padding
self.vis = None
def load_pysot_model(tracker_type):
configpath = "./week3/kalman/pysot/experiments/" + PYSOT_TRACKERS[tracker_type] + \
"/config.yaml"
modelpath = "./week3/kalman/pysot/models/" + PYSOT_TRACKERS[
tracker_type] + ".pth"
cfg.merge_from_file(configpath)
cfg.CUDA = torch.cuda.is_available()
device = torch.device('cuda' if cfg.CUDA else 'cpu')
# load model
model = ModelBuilder()
model.load_state_dict(
torch.load(modelpath, map_location=lambda storage, loc: storage.cpu()))
model.eval().to(device)
return load_pretrain(model, modelpath).cuda().eval()
def __init__(self,dataset=''):
if 'OTB' in dataset:
cfg_file = os.path.join(project_path_,'pysot/experiments/siamrpn_r50_l234_dwxcorr_otb/config.yaml')
snapshot = os.path.join(project_path_,'pysot/experiments/siamrpn_r50_l234_dwxcorr_otb/model.pth')
elif 'LT' in dataset:
cfg_file = os.path.join(project_path_, 'pysot/experiments/siamrpn_r50_l234_dwxcorr_lt/config.yaml')
snapshot = os.path.join(project_path_, 'pysot/experiments/siamrpn_r50_l234_dwxcorr_lt/model.pth')
else:
cfg_file = os.path.join(project_path_, 'pysot/experiments/siamrpn_r50_l234_dwxcorr/config.yaml')
snapshot = os.path.join(project_path_, 'pysot/experiments/siamrpn_r50_l234_dwxcorr/model.pth')
# load config
cfg.merge_from_file(cfg_file)
# create model
self.model = ModelBuilder()# A Neural Network.(a torch.nn.Module)
# load model
self.model = load_pretrain(self.model, snapshot).cuda().eval()
def main():
# load config
cfg.merge_from_file(args.config)
cfg.CUDA = torch.cuda.is_available()
device = torch.device('cuda' if cfg.CUDA else 'cpu')
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).eval().to(device)
# build tracker
tracker = SiamAPNTracker(model, cfg.TRACK)
hp = {'lr': 0.3, 'penalty_k': 0.04, 'window_lr': 0.4}
first_frame = True
if args.video_name:
video_name = args.video_name.split('/')[-1].split('.')[0]
else:
video_name = 'webcam'
cv2.namedWindow(video_name, cv2.WND_PROP_FULLSCREEN)
for frame in get_frames(args.video_name):
if first_frame:
try:
init_rect = cv2.selectROI(video_name, frame, False, False)
except:
exit()
tracker.init(frame, init_rect)
first_frame = False
else:
outputs = tracker.track(frame, hp)
bbox = list(map(int, outputs['bbox']))
cv2.rectangle(frame, (bbox[0], bbox[1]),
(bbox[0] + bbox[2], bbox[1] + bbox[3]), (0, 255, 0),
3)
cv2.imshow(video_name, frame)
cv2.waitKey(40)
def save_siamese_rpn():
# load config
cfg.merge_from_file(args.config)
cfg.BACKBONE.TYPE = 'alexnetlegacy'
# create model
model_legacy = ModelBuilder()
# load model
model_legacy = load_pretrain(model_legacy, args.snapshot).cuda().eval()
cfg.BACKBONE.TYPE = 'alexnet'
# create model
model_alexnet = ModelBuilder()
#
# for key ,item in model.named_parameters():
# print(key,item.shape)
for key, item in model_alexnet.named_parameters():
print(key, item.shape)
name_map = {}
model_legacy_dict = model_legacy.state_dict()
model_alexnet_dict = model_alexnet.state_dict()
for para1, para2 in zip(model_legacy.named_parameters(),
model_alexnet.named_parameters()):
# print(para1[0],para1[1].shape)
print(para1[0])
print(para2[0])
print(para1[1].shape)
print(para2[1].shape)
print("--" * 40)
# print("['{}'--->'{}']".format(para1[0], para2[0]),para1[1].shape, para2[1].shape)
name_map[para1[0]] = para2[0]
print(name_map)
for key, val in name_map.items():
model_alexnet_dict[val] = model_legacy_dict[key]
torch.save(model_alexnet_dict, "siamese_alexnet_rpn.pth")
def train(video, v_idx, attack_region, template_dir):
n_epochs = args.epochs
epsilon = args.epsilon
lr = args.lr
track_model = ModelBuilder()
track_model = load_pretrain(track_model, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker(track_model)
# h, w, _ = video[0][0].shape
attacker = ModelAttacker(args.batch, args.epsilon).cuda().train()
optimizer = optim.Adam(attacker.parameters(), lr=lr)
# optimizer = optim.SGD(attacker.parameters(), lr=lr, momentum=0.9)
for name, param in tracker.model.named_parameters():
param.requires_grad_(False)
# load pretrained
start_epoch = 0
checkpoint_dir = os.path.join(args.savedir, args.dataset, 'checkpoint',
attack_region, video.name)
if os.path.exists(checkpoint_dir):
entries = os.listdir(checkpoint_dir)
if len(entries) > 0:
entries.sort()
start_epoch = int(
entries[-1][-7:-4]) if entries[-1][-7:-4].isnumeric() else 0
if start_epoch == args.epochs:
return
if start_epoch > 0:
state = torch.load(os.path.join(checkpoint_dir, entries[-1]))
attacker.load_state_dict(state['attacker'])
optimizer.load_state_dict(state['optimizer'])
# elif attack_region == 'search':
# checkpoint_dir = os.path.join(args.savedir, args.dataset, 'checkpoint', 'template', video.name)
# assert os.path.exists(
# os.path.join(checkpoint_dir, 'checkpoint_100.pth')), ' missing ' + checkpoint_dir + ' in ' + video.name
# state = torch.load(os.path.join(checkpoint_dir, 'checkpoint_100.pth'))
# attacker.load_state_dict(state['attacker'])
# optimizer.load_state_dict(state['optimizer'])
# generate cropping offset
if attack_region == 'template':
tracker.generate_transition(0, len(video))
elif attack_region == 'search':
tracker.generate_transition(0, len(video))
# disable gradient
if attack_region == 'template':
attacker.adv_x.requires_grad = False
elif attack_region == 'search':
attacker.adv_z.requires_grad = False
training_data = MyDataset()
num_frames = len(video)
it = math.ceil((num_frames - 1) / args.batch)
params = {'batch_size': args.batch, 'shuffle': False, 'num_workers': 6}
for i in range(0, it - 1):
for j in range(0, args.batch):
indx = i * args.batch + j + 1
training_data.add([video[indx][0], video[indx][1]])
for j in range(args.batch * (it - 1), num_frames - 1):
training_data.add([video[j + 1][0], video[j + 1][1]])
img_names = [
x.replace(args.dataset_dir, args.fabricated_dir)
for x in video.img_names
]
del img_names[0]
data_loader = torch.utils.data.DataLoader(training_data, **params)
toc = 0
for epoch in range(0, args.epochs):
if epoch < start_epoch:
continue
# initial frame
img, gt_bbox = video[0]
if attack_region == 'search':
# img_name = video.img_names[0].replace(args.dataset_dir, args.fabricated_dir)
img_name = os.path.join(args.savedir, args.dataset, video.name,
'000099.jpg')
img = cv2.imread(img_name)
gt_bbox, gt_bbox_ = gt_bbox_adaptor(gt_bbox)
state = {'img': img, 'gt_bbox': gt_bbox_, 'video_name': video.name}
state, loss = adversarial_train(0, state, attacker, tracker, optimizer,
gt_bbox_, attack_region, template_dir,
epoch)
pbar = tqdm(enumerate(data_loader), position=0, leave=True)
_loss = []
if attack_region == 'template':
adv_z = []
for (_idx, (idx, (imgs, gt_bboxes))) in enumerate(pbar):
if len(gt_bboxes[0]) == 4:
gt_bboxes = (gt_bboxes[:, 0], gt_bboxes[:, 1], gt_bboxes[:, 0],
gt_bboxes[:, 1] + gt_bboxes[:, 3] - 1,
gt_bboxes[:, 0] + gt_bboxes[:, 2] - 1,
gt_bboxes[:, 1] + gt_bboxes[:, 3] - 1,
gt_bboxes[:, 0] + gt_bboxes[:, 2] - 1,
gt_bboxes[:, 1])
gt_bboxes = torch.stack(gt_bboxes).float()
cx, cy, w, h = get_axis_aligned_bbox_tensor(gt_bboxes)
gt_bboxes_ = torch.stack([cx, cy, w, h])
tic = cv2.getTickCount()
state['img'] = imgs
state, loss = adversarial_train(args.batch * idx + 1, state,
attacker, tracker, optimizer,
gt_bboxes_, attack_region,
template_dir, epoch)
toc += cv2.getTickCount() - tic
if idx < 60:
if idx > 0:
_loss.append(loss)
# pbar.set_postfix_str('%d. Video: %s epoch: %d total %.3f %.3f %.3f %.3f %.3f' %
# (v_idx + 1, video.name, epoch + 1, loss[0], loss[1], loss[2], loss[3],
# attacker.adv_z.mean()))
pbar.set_postfix_str('Video(%d): %s epoch: %d ' %
(v_idx + 1, video.name, epoch + 1))
# pbar.set_postfix_str('%d. Video: %s epoch: %d total %.3f %.3f %.3f %.3f' %
# (v_idx + 1, video.name, epoch + 1, loss[0], loss[1], loss[2], loss[3]))
if attack_region == 'search':
fabricated_dir = '/'.join(img_names[0].split('/')[:-1])
if not os.path.exists(fabricated_dir):
os.makedirs(os.path.join(fabricated_dir))
for i in range(len(imgs)):
x_adv = attacker.add_noise(tracker.x_crops[i],
attacker.adv_x[i])
x_adv = x_adv.unsqueeze(0)
save(imgs[i].data.cpu().numpy(),
x_adv,
state['s_x'],
gt_bboxes_[:, i],
img_names[args.batch * idx + i],
shift=tracker.shift[:, args.batch * idx + i +
1].numpy(),
region=attack_region,
save=True)
toc /= cv2.getTickFrequency()
if attack_region == 'template':
z_adv = attacker.add_noise(tracker.z_crop, attacker.adv_z, epsilon)
img_dir = os.path.join(args.savedir, args.dataset,
state['video_name'])
if not os.path.exists(img_dir):
os.makedirs(os.path.join(img_dir))
save(state['zimg'],
z_adv,
state['sz'],
state['init_gt'],
attack_region,
os.path.join(img_dir,
str(epoch).zfill(6) + '.jpg'),
shift=None,
region=attack_region,
save=True)
_loss = np.asarray(_loss)
_loss_v = sum(_loss, 0) / _loss.shape[0]
pbar.clear()
print('%d. Video: %s Time: %.2fs epoch: %d total %.3f %.3f %.3f' %
(v_idx + 1, video.name, toc, epoch + 1, _loss_v[0], _loss_v[1],
_loss_v[2]))
# save state dict
state_dict = {
'attacker': attacker.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch + 1
}
checkpoint_path = os.path.join(args.savedir, args.dataset,
'checkpoint', attack_region, video.name)
if not os.path.exists(checkpoint_path):
os.makedirs(checkpoint_path)
torch.save(state_dict, os.path.join(checkpoint_path, 'checkpoint.pth'))
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():
rank, world_size = dist_init()
logger.info("init done")
# load cfg
cfg.merge_from_file(args.cfg)
if rank == 0:
if not os.path.exists(cfg.TRAIN.LOG_DIR):
os.makedirs(cfg.TRAIN.LOG_DIR)
init_log('global', logging.INFO)
if cfg.TRAIN.LOG_DIR:
add_file_handler('global',
os.path.join(cfg.TRAIN.LOG_DIR, 'logs.txt'),
logging.INFO)
# logger.info("Version Information: \n{}\n".format(commit()))
logger.info("config \n{}".format(json.dumps(cfg, indent=4)))
# create model
# model = ModelBuilder().cuda().train()
model2 = SiamNet_D().cuda().train() ###
# tracker = build_tracker(model)
# dist_model = DistModule(model)
dist_model2 = DistModule(model2)
# load pretrained backbone weights
if cfg.BACKBONE.PRETRAINED:
cur_path = os.path.dirname(os.path.realpath(__file__))
backbone_path = os.path.join(cur_path, '../', cfg.BACKBONE.PRETRAINED)
load_pretrain(model2.netG.backbone, backbone_path)
# create tensorboard writer
if rank == 0 and cfg.TRAIN.LOG_DIR:
tb_writer = SummaryWriter(cfg.TRAIN.LOG_DIR)
else:
tb_writer = None
# build dataset loader
train_loader = build_data_loader()
# build optimizer and lr_scheduler
# optimizer, lr_scheduler, paramsG = build_opt_lr(dist_model2, cfg.TRAIN.START_EPOCH)
optimizer2, lr_scheduler2 = build_opt_lr2(dist_model2,
cfg.TRAIN.START_EPOCH)
# print("gen:",paramsG)
# print("disc",paramsD)
# resume training
# if cfg.TRAIN.RESUME:
# logger.info("resume from {}".format(cfg.TRAIN.RESUME))
# assert os.path.isfile(cfg.TRAIN.RESUME), \
# '{} is not a valid file.'.format(cfg.TRAIN.RESUME)
# model2, optimizer2, cfg.TRAIN.START_EPOCH = \
# restore_from(model2.module.netG, optimizer2, cfg.TRAIN.RESUME)
# load pretrain
# if cfg.TRAIN.PRETRAINED:
# load_pretrain(model2.netG, cfg.TRAIN.PRETRAINED)
# dist_model = DistModule(model)
dist_model2 = DistModule(model2)
# logger.info(lr_scheduler)
logger.info("model prepare done")
# start training
train(train_loader, tb_writer, dist_model2, optimizer2, lr_scheduler2)
def test(video, v_idx, model_name, template_dir=None, img_names=None):
# create model
track_model = ModelBuilder()
# load model
track_model = load_pretrain(track_model, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker(track_model)
# set writing video parameters
height, width, channels = video[0][0].shape
out = cv2.VideoWriter(
os.path.join(args.savedir, args.dataset, video.name + '.avi'),
cv2.VideoWriter_fourcc('M', 'J', 'P', 'G'), 15, (width, height))
frame_counter = 0
toc = 0
pred_bboxes_adv = []
adv_z = []
pbar = tqdm(enumerate(video), position=0, leave=True)
for idx, (img, gt_bbox) in pbar:
gt_bbox, gt_bbox_ = gt_bbox_adaptor(gt_bbox)
tic = cv2.getTickCount()
pred_bbox, _lost, frame_counter = stoa_track(idx, frame_counter, img,
gt_bbox, tracker,
template_dir, img_names)
pred_bboxes_adv.append(pred_bbox)
toc += cv2.getTickCount() - tic
if idx > 0:
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)
out.write(img)
# save results
if args.dataset not in ['VOT2016', 'VOT2018', 'VOT2019']:
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_adv:
f.write(','.join([str(i) for i in x]) + '\n')
else:
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_adv:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join([vot_float2str("%.4f", i)
for i in x]) + '\n')
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():
# 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():
'''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()
from pysot.utils.model_load import load_pretrain
from pysot.models.model_builder import ModelBuilder
from pysot.core.config import cfg
from pysot.utils.bbox import get_axis_aligned_bbox
from toolkit.datasets import DatasetFactory
from matplotlib.image import imread
from PIL import Image
from got10k.datasets import GOT10k
from got10k.utils.viz import show_frame
if __name__ == '__main__':
cfg.merge_from_file('/home/sourabhswain/Documents/SiamRPN/config.yaml')
net_path = '/home/sourabhswain/Documents/SiamRPN/model.pth'
model = ModelBuilder()
model = load_pretrain(model, net_path).cuda().eval()
tracker = build_tracker(model)
dataset = GOT10k(root_dir='/home/sourabhswain/Documents/SiamRPN/dataset',
subset='val')
#dataset = DatasetFactory.create_dataset(name='GOT-10k',
# dataset_root='/home/sourabhswain/Documents/SiamRPN/dataset',
# load_img=False)
"""
for v_idx, (video, anno) in enumerate(dataset):
toc = 0
pred_bboxes = []
scores = []
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)
if __name__ == "__main__":
torch.set_num_threads(1)
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id
# 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)
dataset_root = os.path.join("/ssd", args.dataset)
# create model
net = models.__dict__[args.arch](anchors_nums=args.anchor_nums,
cls_type=args.cls_type)
net = load_pretrain(net, args.snapshot)
net.eval()
net = net.cuda()
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
# Eval dataset
# root = os.path.realpath(os.path.join(os.path.dirname(__file__),
# '../testing_dataset'))
root = "/ssd"
root = os.path.join(root, args.dataset)
if 'OTB' in args.dataset:
dataset_eval = OTBDataset(args.dataset, root)
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)
# UAVDTdataset = '/media/li/CA5CF8AE5CF89683/research/UAVDT/'
# UAVDTdataset = '/media/li/DATA/VisDrone2019-SOT/'
UAVDTdataset = args.datasetroot
dataset_root = os.path.join(UAVDTdataset)
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker_f(model)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.snapshot.split('/')[-2].split('.')[0]
torch.cuda.synchronize()
# OPE tracking
for v_idx, video in enumerate(dataset):
#if v_idx>40:
o_path=os.path.join('results_rt_raw_f', args.dataset, model_name)
if not os.path.isdir(o_path):
os.makedirs(o_path)
out_path = os.path.join('results_rt_raw_f', args.dataset, model_name, video.name + '.pkl')
if os.path.isfile(out_path):
print('({:3d}) Video: {:12s} already done!'.format(
v_idx+1, video.name))
continue
if args.video != '':
# test one special video
if video.name != args.video:
continue
toc = 0
pred_bboxes = []
scores = []
track_times = []
input_fidx = []
runtime = []
timestamps = []
last_fidx = None
n_frame=len(video)
t_total = n_frame/args.fps
t_start = perf_counter()
while 1:
t1 = perf_counter()
t_elapsed=t1-t_start
if t_elapsed>t_total:
break
# identify latest available frame
fidx_continous = t_elapsed*args.fps
fidx = int(np.floor(fidx_continous))
#if the tracker finishes current frame before next frame comes, continue
if fidx == last_fidx:
continue
last_fidx=fidx
tic = cv2.getTickCount()
(img,gt_bbox)=video[fidx]
if fidx == 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_)
torch.cuda.synchronize()
t2 = perf_counter()
t_elapsed=t2-t_start
timestamps.append(t_elapsed)
runtime.append(t2-t1)
pred_bbox = gt_bbox_
scores.append(None)
pred_bboxes.append(pred_bbox)
input_fidx.append(fidx)
else:
box_f = tracker.forecaster.forecast(fidx, input_fidx[-1], np.array([pred_bboxes[-1]]))
outputs = tracker.track(img,box_f[0])
torch.cuda.synchronize()
t2 = perf_counter()
t_elapsed=t2-t_start
timestamps.append(t_elapsed)
runtime.append(t2-t1)
pred_bbox = outputs['bbox']
pred_bboxes.append(pred_bbox)
scores.append(outputs['best_score'])
input_fidx.append(fidx)
if t_elapsed>t_total:
break
toc += cv2.getTickCount() - tic
track_times.append((cv2.getTickCount() - tic)/cv2.getTickFrequency())
#save results and run time
if args.overwrite or not os.path.isfile(out_path):
pickle.dump({
'results_raw': pred_bboxes,
'timestamps': timestamps,
'input_fidx': input_fidx,
'runtime': runtime,
}, open(out_path, 'wb'))
toc /= cv2.getTickFrequency()
# save results
# 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, fidx / toc))
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
# 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)
cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(cur_dir, '../test_dataset', args.dataset)
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
tracker = SiamAPNTracker(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] + str(cfg.TRACK.w1)
# 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
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():
mode = args.mode
# 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_model = ModelBuilder()
track_model0 = ModelBuilder()
lr = args.lr
# load model
track_model = load_pretrain(track_model, args.snapshot).cuda().eval()
track_model0 = load_pretrain(track_model0, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker(track_model)
tracker0 = build_tracker(track_model0)
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 tracker.model.named_parameters():
param.requires_grad_(False)
# for name, param in attacker.named_parameters():
# if 'backbone' in name or 'neck' in name or 'rpn_head' in name:
# 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
pred_bboxes = []
pred_bboxes_adv = []
pbar = tqdm(enumerate(video))
for idx, (img, gt_bbox) in pbar:
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]
tic = cv2.getTickCount()
##########################################
# # # for state of the art tracking # # #
##########################################
if mode == 0:
pred_bbox, _lost, frame_counter = stoa_track(
idx, frame_counter, img, gt_bbox, tracker0)
##########################################
# # # # # adversarial tracking # # # # #
##########################################
if idx == 0:
state = {
'img': img,
'gt_bbox': gt_bbox,
'video_name': video.name
}
else:
state['img'] = img
if mode == 1:
optimizer, state, ad_bbox = \
adversarial_train(idx, frame_counter_adv, state, attacker, tracker, optimizer, pbar)
if idx == 0:
break
toc += cv2.getTickCount() - tic
if idx > 0 and mode == 0:
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)
if idx > 0 and mode == 1:
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)
__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)
out.write(img)
toc /= cv2.getTickFrequency()
if mode == 0:
# save results
if args.dataset == 'OTB100':
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_adv:
f.write(','.join([str(i) for i in x]) + '\n')
else:
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))
# ii = 0
# with open(result_path, 'r') as f:
# xs = f.readlines()
# for x in xs:
# if ii == 0:
# pred_bboxes_adv[0] = ','.join([vot_float2str("%.4f", i) for i in pred_bboxes_adv[0]]) + '\n'
# else:
# pred_bboxes_adv.append(x)
# ii += 1
#
# with open(result_path, 'w') as f:
# for x in pred_bboxes_adv:
# f.write(x)
with open(result_path, 'w') as f:
for x in pred_bboxes_adv:
if isinstance(x, int):
f.write("{:d}\n".format(x))
else:
f.write(','.join(
[vot_float2str("%.4f", i)
for i in x]) + '\n')
raise NotImplementedError
if __name__ == '__main__':
torch.backends.cudnn.benchmark = True
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 = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
benchmark = EAOBenchmark(dataset)
model = ModelBuilder()
model = load_pretrain(model, args.snapshot).cuda().eval()
# the resources you computer have, object_store_memory is shm
ray.init(num_gpus=1, num_cpus=8, object_store_memory=30000000000)
tune.register_trainable('fitness', fitness)
# define search space
params = {
'penalty_k': hp.quniform('penalty_k', 0.001, 0.6, 0.001),
'lr': hp.quniform('scale_lr', 0.1, 0.8, 0.001),
'window_influence': hp.quniform('window_influence', 0.05, 0.65, 0.001),
'search_region': hp.choice('search_region', [255]),
}
# stop condition for VOT and OTB
if args.dataset.startswith('VOT'):
def main():
# load config
cfg.merge_from_file(args.config)
# hp_search
params = getattr(cfg.HP_SEARCH, args.dataset)
hp = {'lr': params[0], 'penalty_k': params[1], 'window_lr': params[2]}
cur_dir = os.path.dirname(os.path.realpath(__file__))
dataset_root = os.path.join(cur_dir, '../testing_dataset', args.dataset)
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
tracker = SiamCARTracker(model, cfg.TRACK)
# create dataset
dataset = DatasetFactory.create_dataset(name=args.dataset,
dataset_root=dataset_root,
load_img=False)
model_name = args.snapshot.split('/')[-2] + str(hp['lr']) + '_' + str(
hp['penalty_k']) + '_' + str(hp['window_lr'])
# 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:
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(pred_bbox)
else:
outputs = tracker.track(img, hp)
pred_bbox = outputs['bbox']
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:
if not any(map(math.isnan, gt_bbox)):
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
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))
os.chdir(model_path)
save_file = '../%s' % dataset
shutil.make_archive(save_file, 'zip')
print('Records saved at', save_file + '.zip')
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():
# load config
cfg.merge_from_file(args.config)
# create model
model = ModelBuilder()
# load model
model = load_pretrain(model, args.snapshot).cuda().eval()
# build tracker
tracker = build_tracker(model)
#model_name = args.snapshot.split('/')[-1].split('.')[0]
#total_lost = 0
#cur_dir = os.path.dirname(os.path.realpath(__file__))
#dataset_root = os.path.join(cur_dir, '../testing_dataset', args.dataset)
video_path = '/home/yuuzhao/Documents/project/pysot/testing_dataset/VOT2016'
#lists = open('/home/lichao/tracking/LaSOT_Evaluation_Toolkit/sequence_evaluation_config/' + setfile + '.txt', 'r')
#list_file = [line.strip() for line in lists]
category = os.listdir(video_path)
category.sort()
# create dataset
#dataset = DatasetFactory.create_dataset(name=args.dataset,dataset_root=dataset_root,load_img=False)
template_acc = []
template_cur = []
init0 = []
init = []
pre = []
gt = [] # init0 is reset init
print("Category & Video:")
for tmp_cat in category:
tmp_cat_path = temp_path + '/' + tmp_cat
if not os.path.isdir(tmp_cat_path):
os.makedirs(tmp_cat_path)
print("Category:", tmp_cat)
video = os.listdir(join(video_path, tmp_cat))
video.sort()
#video_cut = video[0:frames_of_each_video]
frame = 0
#for picture in video_cut: # 这个循环或许该去掉
# print("Frame:", picture)
gt_path = join(video_path, tmp_cat, 'groundtruth.txt')
ground_truth = np.loadtxt(gt_path, delimiter=',')
# num_frames = len(ground_truth); # num_frames = min(num_frames, frame_max)
num_frames = frames_of_each_video
# print("num_frames: ",num_frames)
img_path = join(video_path, tmp_cat)
# print("imgpath",img_path)
imgFiles = [
join(img_path, '%08d.jpg') % i for i in range(1, num_frames + 1)
]
while frame < num_frames:
print("frame:", frame)
Polygon = ground_truth[frame]
cx, cy, w, h = get_axis_aligned_bbox(Polygon)
gt_rect = [cx, cy, w, h]
image_file = imgFiles[frame]
# target_pos, target_sz = np.array([cx, cy]), np.array([w, h])
img = cv2.imread(image_file) # HxWxC
if frame == 0:
tracker.init(img, gt_rect)
if w * h != 0:
# image_file = imgFiles[frame]
# target_pos, target_sz = np.array([cx, cy]), np.array([w, h])
# img = cv2.imread(image_file) # HxWxC
zf_acc = tracker.get_zf(img, gt_rect)
output = tracker.track(img)
pre_rect = output['bbox']
zf_pre = tracker.get_zf(img, pre_rect)
template_acc.append(zf_acc)
template_cur.append((zf_pre))
print("ACC&PRE")
init0.append(0)
init.append(frame)
frame_reset = 0
pre.append(0)
gt.append(1)
while frame < (num_frames - 1):
print("while ", frame, "<", num_frames)
frame = frame + 1
frame_reset = frame_reset + 1
image_file = imgFiles[frame]
if not image_file:
break
Polygon = ground_truth[frame]
cx, cy, w, h = get_axis_aligned_bbox(Polygon)
gt_rect = [cx, cy, w, h]
img = cv2.imread(image_file) # HxWxC
zf_acc = tracker.get_zf(img, gt_rect)
output = tracker.track(img)
pre_rect = output['bbox']
zf_pre = tracker.get_zf(img, pre_rect)
# print("zf_pre:",zf_pre.shape)
# print("zf_acc:",zf_acc.shape)
# pdb.set_trace()
template_acc.append(zf_acc)
template_cur.append(zf_pre)
init0.append(frame_reset)
init.append(frame)
pre.append(1)
if frame == (num_frames - 1): # last frame
print("if frame == num_frames-1")
gt.append(0)
else:
gt.append(1)
pre_rect_arr = np.array(pre_rect)
cx, cy, w, h = get_axis_aligned_bbox(pre_rect_arr)
target_pos, target_siz = np.array([cx,
cy]), np.array([w, h])
res = cxy_wh_2_rect(target_pos, target_siz)
if reset:
cx, cy, w, h = get_axis_aligned_bbox(
ground_truth[frame])
gt_rect = [cx, cy, w, h]
gt_rect = np.array(gt_rect)
iou = overlap_ratio(gt_rect, res)
if iou <= 0:
break
else:
print("else")
template_acc.append(
torch.zeros([1, 3, 127, 127], dtype=torch.float32))
template_cur.append(
torch.zeros([1, 3, 127, 127], dtype=torch.float32))
init0.append(0)
init.append(frame)
pre.append(1)
if frame == (num_frames - 1): # last frame
gt.append(0)
else:
gt.append(1)
frame = frame + 1 # skip
#写出一次
#print("template_acc:",template_acc)
#print("template_cur:",template_cur)
#print("init:", init)
#print("init0:",init0)
#print("pre:",pre)
#template_acc_con = np.concatenate(template_acc);
#template_cur_con = np.concatenate(template_cur)
print("write for each video")
np.save(tmp_cat_path + '/template', template_acc)
np.save(tmp_cat_path + '/templatei', template_cur)
np.save(tmp_cat_path + '/init0', init0)
np.save(tmp_cat_path + '/init', init)
np.save(tmp_cat_path + '/pre', pre)
np.save(tmp_cat_path + '/gt', gt)
print("template")