def eval_det(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
for path, img, img0 in dataloader:
if frame_id < 302:
frame_id += 1
continue
elif frame_id > 302:
break
else:
print(frame_id)
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# if frame_id>20:
# break
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
dets = tracker.detect(blob, img0)
# print(path, dets)
tlbrs = []
scores = []
class_ids = []
for det in dets:
tlbrs.append(det[:4])
scores.append(det[4])
class_ids.append(int(det[5] - 1))
# print(class_ids)
if show_image or save_dir is not None:
online_im = vis.plot_detections(img0,
tlbrs,
scores=None,
ids=class_ids)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30,
use_cuda=True):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
#for path, img, img0 in dataloader:
for i, (path, img, img0) in enumerate(dataloader):
#if i % 8 != 0:
#continue
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
if use_cuda:
blob = torch.from_numpy(img).cuda().unsqueeze(0)
else:
blob = torch.from_numpy(img).unsqueeze(0)
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
#online_scores = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
#online_scores.append(t.score)
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids))
#results.append((frame_id + 1, online_tlwhs, online_ids, online_scores))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
write_results(result_filename, results, data_type)
#write_results_score(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval(self, skip_frame=1, show_image=False):
tracker = JDETracker(self.opt, frame_rate=self.frame_rate)
timer = Timer()
frame_id = 0
for i, (path, img, img0) in enumerate(self.dataloader):
if i % skip_frame != 0:
continue
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
if self.use_cuda:
blob = torch.from_numpy(img).cuda().unsqueeze(0)
else:
blob = torch.from_numpy(img).unsqueeze(0)
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
online_scores = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > self.opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
online_scores.append(t.score)
timer.toc()
tmp_result = {
"frame_id": frame_id + 1,
"bounding_box": online_tlwhs,
"ids": online_ids,
"scores": online_scores
}
self.send_result(tmp_result, raw_img=img0)
frame_id += 1
if show_image:
online_im = self.send_image(img0, online_tlwhs, online_ids,
frame_id,
1. / max(1e-5, timer.average_time))
cv2.imshow('Result', online_im)
if self.video_saver is not None:
self.video_saver.release()
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
save_dir=None,
show_image=True,
frame_rate=30):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
writer = VideoWriter(save_dir, dataloader)
for path, img, img0 in dataloader:
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
# cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)), online_im)
writer.write(online_im)
frame_id += 1
print("***************************** DONE *****************************")
def eval_seq(opt, save_dir='', frame_rate=30):
dataloader = LoadImages(opt.multi_test_images_dataset)
opt = opt.update_dataset_info_and_set_heads(opt, dataloader)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
fd = os.open(save_dir + '/args.txt', os.O_RDWR)
for i, (path, img, img0) in enumerate(dataloader):
# run tracking
timer.tic()
blob = torch.from_numpy(img).unsqueeze(0)
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
online_tlwhs.append(tlwh)
online_ids.append(tid)
centre = tlwh[:2] + tlwh[2:] / 2
text = str(frame_id) + ' ' + str(tid) + ' ' + str(
centre[1]) + ' ' + str(centre[0]) + '\n'
os.write(fd, str.encode(text))
timer.toc()
# save results
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
os.close(fd)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30,
use_cuda=True,
gt_path=None):
gt_results_dict = dict()
if gt_path:
if os.path.isfile(gt_path):
with open(gt_path, 'r') as f:
for line in f.readlines():
linelist = line.split(',')
if len(linelist) < 7:
continue
fid = int(linelist[0])
if fid < 1:
continue
gt_results_dict.setdefault(fid, list())
box_size = float(linelist[4]) * float(linelist[5])
score = 1
# if box_size > 7000:
# if box_size <= 7000 or box_size >= 15000:
# if box_size < 15000:
# continue
tlwh = tuple(map(float, linelist[2:6]))
target_id = (linelist[1])
gt_results_dict[fid].append((tlwh, target_id, score))
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
#for path, img, img0 in dataloader:
for i, (path, img, img0) in enumerate(dataloader):
#if i % 8 != 0:
#continue
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
if use_cuda:
blob = torch.from_numpy(img).cuda().unsqueeze(0)
else:
blob = torch.from_numpy(img).unsqueeze(0)
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
#online_scores = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
#online_scores.append(t.score)
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids))
#results.append((frame_id + 1, online_tlwhs, online_ids, online_scores))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time,
gt_box=gt_results_dict[frame_id + 1])
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
write_results(result_filename, results, data_type)
#write_results_score(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
'''
img: Normalized RGB image
img0: BGR image
'''
for path, img, img0 in dataloader:
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
# print("\n==> blob.size", blob.size()) 1, 3, 608, 1088
'''
tracker update
'''
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
'''
print("==> [track.eval_seq] tracker's output-> online_targets:", online_targets)
try:
print("==> [track.eval_seq] len(online_tlwhs):", len(online_tlwhs))
print("==> [track.eval_seq] online_tlwhs[0]:", online_tlwhs[0])
print("==> [track.eval_seq] online_ids[0]:", online_ids[0])
except:
pass
partial output:
==> [multi-tracker.update] len(output_stracks): 5
==> [track.eval_seq] tracker's output-> online_targets: [OT_1_(1-13), OT_2_(1-13), OT_3_(1-13), OT_20_(10-13), OT_7_(2-13)]
==> [track.eval_seq] len(online_tlwhs): 5
==> [track.eval_seq] online_tlwhs[0]: [ 802.38 163.64 24.074 57.376]
==> [track.eval_seq] online_ids[0]: 1
==> [multi-tracker.update] len(output_stracks): 7
==> [track.eval_seq] tracker's output-> online_targets: [OT_1_(1-14), OT_2_(1-14), OT_3_(1-14), OT_20_(10-14), OT_7_(2-14), OT_23_(13-14), OT_13_(4-14)]
==> [track.eval_seq] len(online_tlwhs): 7
==> [track.eval_seq] online_tlwhs[0]: [ 809.96 163.69 25.305 60.319]
==> [track.eval_seq] online_ids[0]: 1
==> [multi-tracker.update] len(output_stracks): 7
==> [track.eval_seq] tracker's output-> online_targets: [OT_1_(1-15), OT_2_(1-15), OT_3_(1-15), OT_20_(10-15), OT_7_(2-15), OT_23_(13-15), OT_19_(10-15)]
==> [track.eval_seq] len(online_tlwhs): 7
==> [track.eval_seq] online_tlwhs[0]: [ 818.46 164.4 26.832 63.971]
==> [track.eval_seq] online_ids[0]: 1
'''
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
write_results(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
len_all = len(dataloader)
start_frame = int(len_all / 2)
frame_id = int(len_all / 2)
out_queue = Deque(maxlen=5)
for i, (path, img, img0) in enumerate(dataloader):
if i < start_frame:
continue
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
while len(out_queue) < out_queue.maxlen - 1:
output = tracker.model.features(blob)[-1]
out_queue.append(output)
ref = torch.cat(list(out_queue), dim=0)
output = tracker.model((blob, ref))[0]
out_queue.append(output.pop('raw'))
online_targets = tracker.update(blob, img0, output)
online_tlwhs = []
online_ids = []
#online_scores = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
#online_scores.append(t.score)
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids))
#results.append((frame_id + 1, online_tlwhs, online_ids, online_scores))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
write_results(result_filename, results, data_type)
#write_results_score(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
tid_max, tid_temp = 1, 1
for path, img, img0 in dataloader:
if frame_id < 300:
frame_id += 1
continue
elif frame_id > 302:
break
else:
print(frame_id)
filename = path.split('/')[-1]
if '0000001' in path:
tid_max = tid_temp + 1
print(path, tid_max)
tracker = JDETracker(opt, frame_rate=frame_rate)
if frame_id % 100 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# if frame_id >20:
# break
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
online_targets = tracker.update_sep3(
blob,
img0,
conf_thres=[0.4, 0.3, 0.4, 0.4, 0.4, 0.4, 0.4, 0.5, 0.5])
# print(online_targets)
online_tlwhs = []
online_ids = []
online_cids = [] #class id
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id + tid_max
tcid = t.class_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area: # and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
online_cids.append(tcid)
tid_temp = max(tid, tid_temp)
timer.toc()
# save results
results.append((filename, online_tlwhs, online_ids, online_cids))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
write_results(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir='.',
show_image=True,
frame_rate=25):
tracker = JDETracker(opt, frame_rate=frame_rate)
p = path_root_index[5]
if save_dir:
save_dir = osp.join(save_dir, p)
mkdir_if_missing(save_dir)
image_path = getimage_path(path_root + p)
timer = Timer()
results = []
frame_id = -1
result_array_list = []
result = []
for path in image_path:
# img=cv2.imread(path)
img0 = cv2.imread(path) # BGR
# assert img0 is not None, 'Failed to load ' + img_path
img_height = img0.shape[0]
img_width = img0.shape[1]
# print(img_height,img_width)
# print(img0.shape)
# Padded resize
img, _, _, _ = letterbox(img0, height=608, width=1088)
# Normalize RGB
img = img[:, :, ::-1].transpose(2, 0, 1)
img = np.ascontiguousarray(img, dtype=np.float32)
img /= 255.0
frame_id += 1
# if frame_id % 20 == 0:
# logger.info('Processing frame {} ({:.2f} fps)'.format(frame_id, 1. / max(1e-5, timer.average_time)))
# if frame_id==2:
# break
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
online_cref = []
# result_array_list=[]
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
confidence = t.score
vertical = tlwh[2] / tlwh[3] > 1.6
if confidence < 0.3:
if tlwh[2] * tlwh[3] > 2700 and not vertical and tlwh[
2] * tlwh[3] < 100000:
res = [frame_id, tid]
res += list(tlwh)
res += [1, 0]
online_tlwhs.append(tlwh)
result_array_list.append(res)
online_cref.append(confidence)
# print(confidence)
# result_array_list.append(tlwh)
online_ids.append(tid)
elif confidence >= 0.3:
if tlwh[2] * tlwh[3] > 1000 and not vertical and tlwh[
2] * tlwh[3] < 100000:
# print(tlwh[2] * tlwh[3])
res = [frame_id, tid]
res += list(tlwh)
res += [1, 0]
online_tlwhs.append(tlwh)
result_array_list.append(res)
online_cref.append(confidence)
# print(confidence)
# result_array_list.append(tlwh)
online_ids.append(tid)
# if frame_id==2:
# break
timer.toc()
# save results
print(frame_id)
# if result_array_list:
online_tlwhs = np.array(online_tlwhs)
online_ids = np.array(online_ids)
online_cref = np.array(online_cref)
# print(online_tlwhs)
# print(online_tlwhs.shape)
# print(online_ids.shape)
# pick=non_max_suppression(online_tlwhs,0.7,online_cref)
# online_tlwhsnms=online_tlwhs[pick]
# online_idsnms=online_ids[pick]
# online_crefnms=online_cref[pick]
# result_array_list2=np.array(result_array_list).copy()[pick]
# result+=list(result_array_list2)
# print(result)
# print(frame_id,online_idsnms)
# result.append(online_tlwhsnms)
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
scores=online_cref,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
cv2.waitKey(1)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
res_array = np.array(result_array_list)
# print(res_array)
# print(res_array.shape)
tmp_data1 = np.where(res_array[:, [2]] < 0,
res_array[:, [2]] + res_array[:, [4]], res_array[:,
[4]])
res_array[:, [4]] = tmp_data1
tmp_data = np.where(res_array[:, [3]] < 0,
res_array[:, [3]] + res_array[:, [5]], res_array[:,
[5]])
res_array[:, [5]] = tmp_data
res_array[:, [2, 3]] = np.maximum(res_array[:, [2, 3]], 0)
# print(res_array)
res_array = np.round(res_array, 0)
# res_array=cutmorecord(res_array,img_width,img_height)
# print(res_array)
np.savetxt("{}.txt".format(p),
res_array,
fmt='%d,%d,%d,%d,%d,%d,%d,%d',
delimiter=',')
# save results
# write_results(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
polygon,
paths,
data_type,
result_filename,
frame_dir=None,
save_dir=None,
bbox_dir=None,
show_image=True,
frame_rate=30):
count = 0
if save_dir:
mkdir_if_missing(save_dir)
if bbox_dir:
mkdir_if_missing(bbox_dir)
if frame_dir:
mkdir_if_missing(frame_dir)
tracker = JDETracker(opt, polygon, paths, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 1
f = open(opt.input_video.split('/')[-1][:-4] + '.txt', 'w')
for path, img, img0 in dataloader:
img0_clone = copy.copy(img0)
if frame_id % 1 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(
0) if opt.gpus[0] >= 0 else torch.from_numpy(img).cpu().unsqueeze(
0)
online_targets, detection_boxes, out_of_polygon_tracklet = tracker.update(
blob, img0)
if len(out_of_polygon_tracklet) > 0:
for track in np.asarray(out_of_polygon_tracklet)[:, 2]:
if track in ['person', 'bicycle', 'motorcycle']:
count += 1
print('count : ' + str(count))
online_tlwhs = []
online_ids = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
if tlwh[2] * tlwh[3] > opt.min_box_area:
online_tlwhs.append(tlwh)
online_ids.append(tid)
#bbox detection plot
box_tlbrs = []
box_scores = []
box_classes = []
box_occlusions = []
img_bbox = img0.copy()
for box in detection_boxes:
tlbr = box.tlbr
tlwh = box.tlwh
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area:
box_tlbrs.append(tlbr)
box_scores.append(box.score)
box_classes.append(box.infer_type())
box_occlusions.append('occ' if box.occlusion_status ==
True else 'non_occ')
timer.toc()
# save results
for track in out_of_polygon_tracklet:
frame_idx, id, classes, movement = track
results.append((opt.input_video.split('/')[-1][:-4], frame_idx,
classes, movement))
f.write(','.join([
opt.input_video.split('/')[-1][:-4],
str(frame_idx),
str(classes),
str(movement)
]) + '\n')
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time,
out_track=out_of_polygon_tracklet)
bbox_im = vis.plot_detections(img_bbox,
box_tlbrs,
scores=box_scores,
box_occlusion=None,
btypes=box_classes)
if show_image:
cv2.polylines(online_im, [np.asarray(polygon)], True,
(0, 255, 255))
cv2.polylines(bbox_im, [np.asarray(polygon)], True, (0, 255, 255))
cv2.polylines(img0_clone, [np.asarray(polygon)], True,
(0, 255, 255))
cv2.imshow('online_im', online_im)
cv2.imshow('bbox_im', bbox_im)
if save_dir is not None:
cv2.polylines(online_im, [np.asarray(polygon)], True,
(0, 255, 255))
cv2.polylines(bbox_im, [np.asarray(polygon)], True, (0, 255, 255))
cv2.polylines(img0_clone, [np.asarray(polygon)], True,
(0, 255, 255))
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
cv2.imwrite(os.path.join(bbox_dir, '{:05d}.jpg'.format(frame_id)),
bbox_im)
cv2.imwrite(os.path.join(frame_dir, '{:05d}.jpg'.format(frame_id)),
img0_clone)
frame_id += 1
# save results
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30):
#verifica che sia presente la cartella di salvataggio e in caso non lo sia la crea
if save_dir:
mkdir_if_missing(save_dir)
#crea il JDETracker che esegue le seguenti operazioni (che vengono trovate in src/lib/tracker/multitracker.py nella funzione init)
#1) Verifica se il programma va eseguito con CPU o GPU
#2) Crea il modello e lo valuta
#3) Applica un Kalman Filter: e’ uno strumento per stimare lo stato di un sistema dinamico lineare perturbato da rumore, sulla base di misure (o osservazioni) linearmente dipendenti
tracker = JDETracker(opt, frame_rate=frame_rate)
#viene inizializzato il timer per monitorare il tempo di elaborazione
timer = Timer()
#inizializzazione array dei risultati
results = []
#identificatore del frame
frame_id = 0
#scorriamo gli elementi all'interno del dataloader che erano le immagini del database caricate nel main
for path, img, img0 in dataloader:
#visualizza il frame rate dopo 20 frame elaborati
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
#questa funzione è specifica per le GPU NVIDIA, infatti usa CUDA, si può sostituire con CPU nel caso si voglia eseguire il programma con CPU
#crea una matrice multidimensionale di almeno 3 elementi per rappresentare l'immagine img
blob = torch.from_numpy(img).cuda().unsqueeze(0)
#viene applicata la funzione update del JDETracker che svolge le seguenti funzioni (src/lib/tracker/multitracker.py)
#1) Vengono passati come parametri gli elementi blob e img0
#2) Da entrambi si estraggono altezza e larghezza e vengono memorizzate in una variabile
#3) viene creata una variabile c che memorizza un array di float che contiene come due elementi altezza e larghezza dimezzata di img0
#4) viene creata una variabile s che memorizza il massimo fra max(float(inp_width) (blob) / float(inp_height) (blob) * height (img0), width (img0)) * 1.0
#5) Network forward, get detections & embeddings: rileva gli oggetti e gli aggiunge a una pila di tracklet
#6) Prima associazione con Embedding: Da Paper, il ruolo dell'embedding è quello di distinguere diversi tipi di oggetti
#7) Seconda associazione con IoU
#8) Inizializza nuovi Stracks
#9) Aggiorna lo stato
#10) Ritorna nella variabile il valore degli stracks attivi
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
#vengono iterati i vari stracks
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
write_results(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq_realtime(opt, save_dir=None, show_image=True, frame_rate=30):
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
frame_id = 0
cv2.namedWindow('online_im', cv2.WINDOW_FREERATIO)
cv2.setWindowProperty('online_im', cv2.WND_PROP_AUTOSIZE,
cv2.WND_PROP_AUTOSIZE)
cap = cv2.VideoCapture(0)
ret, im = cap.read()
height = im.shape[0]
width = im.shape[1]
while True:
ret, img0 = cap.read()
# Padded resize
img, _, _, _ = letterbox(img0, height=height, width=width)
# Normalize RGB
img = img[:, :, ::-1].transpose(2, 0, 1)
img = np.ascontiguousarray(img, dtype=np.float32)
img /= 255.0
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
timer.toc()
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
text_scale = max(1, online_im.shape[1] / 1600.)
text_thickness = 1 if text_scale > 1.1 else 1
line_thickness = max(1, int(online_im.shape[1] / 500.))
cv2.putText(online_im,
'Press ESC to STOP', (300, int(15 * text_scale)),
cv2.FONT_HERSHEY_PLAIN,
text_scale, (0, 0, 255),
thickness=2)
cv2.imshow('online_im', online_im)
key = cv2.waitKey(1)
if key == 27:
break
frame_id += 1
cv2.destroyAllWindows()
def main(opt):
acl_resource = AclResource()
acl_resource.init()
mot_model = Model('../model/dlav0.om')
# Create output dir if not exist; default outputs
result_root = opt.output_root if opt.output_root != '' else '.'
mkdir_if_missing(result_root)
video_name = os.path.basename(opt.input_video).replace(' ',
'_').split('.')[0]
# setup dataloader, use LoadVideo or LoadImages
dataloader = LoadVideo(opt.input_video, (1088, 608))
# result_filename = os.path.join(result_root, 'results.txt')
frame_rate = dataloader.frame_rate
# dir for output images; default: outputs/'VideoFileName'
save_dir = os.path.join(result_root, video_name)
if save_dir and os.path.exists(save_dir) and opt.rm_prev:
shutil.rmtree(save_dir)
mkdir_if_missing(save_dir)
# initialize tracker
tracker = JDETracker(opt, mot_model, frame_rate=frame_rate)
timer = Timer()
results = []
# img: h w c; 608 1088 3
# img0: c h w; 3 608 1088
for frame_id, (path, img, img0) in enumerate(dataloader):
if frame_id % 20 == 0:
print('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking, start tracking timer
timer.tic()
# list of Tracklet; see multitracker.STrack
online_targets = tracker.update(np.array([img]), img0)
# prepare for drawing, get all bbox and id
online_tlwhs = []
online_ids = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
timer.toc()
# draw bbox and id
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30):
if save_dir:
mkdir_if_missing(save_dir)
if opt.gpus[0] >= 0:
opt.device = 'gpu'
else:
opt.device = 'cpu'
paddle.set_device(opt.device)
print('Creating model...')
model = create_model(opt.arch, opt.heads, opt.head_conv)
model = load_model(model, opt.load_model)
# model = torch.nn.DataParallel(model)
# model = model.to(opt.device)
model.eval()
timer = Timer()
results = []
frame_id = 0
for path, img, img0 in dataloader:
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run detecting
timer.tic()
# blob = torch.from_numpy(img).cuda().unsqueeze(0)
blob = paddle.to_tensor(img).unsqueeze(0)
width = img0.shape[1]
height = img0.shape[0]
inp_height = blob.shape[2]
inp_width = blob.shape[3]
c = np.array([width / 2., height / 2.], dtype=np.float32)
s = max(float(inp_width) / float(inp_height) * height, width) * 1.0
meta = {
'c': c,
's': s,
'out_height': inp_height // opt.down_ratio,
'out_width': inp_width // opt.down_ratio
}
# with torch.no_grad():
with paddle.clear_grad():
output = model(blob)[-1]
hm = output['hm'].sigmoid_()
wh = output['wh']
reg = output['reg'] if opt.reg_offset else None
dets, inds = mot_decode(hm, wh, reg=reg, ltrb=opt.ltrb, K=opt.K)
dets = post_process(opt, dets, meta)
dets = merge_outputs(opt, [dets])[1]
dets = dets[dets[:, 4] > 0.1]
dets[:, :4] = tlbr2tlwh(dets[:, :4])
tlwhs = []
scores = []
for *tlwh, conf in dets:
tlwhs.append(tlwh)
scores.append(conf)
timer.toc()
# save results
results.append((frame_id + 1, tlwhs, scores))
frame_id += 1
# save results
write_results_score(result_filename, results)
#write_results_score_hie(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30,
use_cuda=True):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
seq = '.'.join(osp.basename(result_filename).split('.')[:-1])
gt = TrackSet(
'/mnt/lustre/share/fengweitao/myval/images/val/%s/gt/gt.txt' % seq)
#for path, img, img0 in dataloader:
out_queue = Deque(maxlen=6)
gt_queue = Deque(maxlen=6)
gt_raw_q = Deque(maxlen=6)
for i, (path, img, img0) in enumerate(dataloader):
#if i % 8 != 0:
#continue
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
time_clean()
time_sync('tracker all')
gt_row = gt[frame_id + 1]
if use_cuda:
blob = torch.from_numpy(img).cuda().unsqueeze(0)
else:
blob = torch.from_numpy(img).unsqueeze(0)
with torch.no_grad():
time_sync('forward_half and queue')
out = tracker.model.backend.forward_half(
blob, roi_num=tracker.model.backend.roi_top_k)
out_ref = {}
out_ref.update(out)
out_ref['rois'], out_ref[
'roi_scores'] = tracker.model.backend.get_rois(
out['rpn_map'], 300)
gt_rois = get_gt_rois(out['rois'], gt_row,
out['rpn_map'].shape[-2:], img0.shape[:2])
gt_rois_ref = get_gt_rois(out_ref['rois'], gt_row,
out['rpn_map'].shape[-2:],
img0.shape[:2])
refs = tracker.model.backend.forward_rois(out_ref)
out_queue.append(refs)
gt_queue.append(gt_rois_ref)
gt_raw_q.append(gt_row)
while len(out_queue) < out_queue.maxlen:
out_queue.append(refs)
gt_queue.append(gt_rois_ref)
gt_raw_q.append(gt_row)
time_sync('forward_half and queue')
time_sync('forward_all')
def wrap_gt(drow, img0):
labels = [[
0,
int(d.uid), d.cx / img0.shape[1], d.cy / img0.shape[0],
d.w / img0.shape[1], d.h / img0.shape[0]
] for d in drow]
labels = np.array(labels)
# print(labels)
input_gt = reldata.parse_input_data(blob, opt, labels)
input_gt = to_device(input_gt, device=blob.device)
return input_gt
labels = wrap_gt(gt_row, img0)
# print(labels)
labels_ref = []
for one in gt_raw_q:
labels_ref_one = wrap_gt(one, img0)
labels_ref.append(labels_ref_one)
input_gt = {}
for k in labels:
if k in ['hm', 'reg_mask', 'ind', 'ids']:
input_gt[k] = labels[k], [one[k] for one in labels_ref]
else:
input_gt[k] = labels[k]
output, stuffs, _ = tracker.model.forward_half(out,
out_queue,
inputs=input_gt)
gt_ref_rois = []
for gt_key_i in range(len(gt_queue[0])):
tmp_gt_rois = np.concatenate(
[one[gt_key_i] for one in gt_queue])
gt_ref_rois.append(tmp_gt_rois)
time_sync('forward_all')
# info_debug(output)
# jj = output['rois'][0] % out['rpn_map'].shape[-1]
# ii = output['rois'][0] // out['rpn_map'].shape[-1]
# u = output['rpn_map'][:, :, ii, jj]
# print(output['hm'][:20])
# input()
# import pickle
# print(u.shape)
# print(u.flatten().sigmoid())
# print((u.flatten().sigmoid() - output['hm'].flatten().sigmoid()).abs().max())
# with open('two_none.pkll', 'wb') as fd:
# pickle.dump(output, fd)
# input()
time_sync('tracker update')
online_targets = tracker.update(blob, img0, output)
time_sync('tracker update')
online_tlwhs = []
online_ids = []
#online_scores = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
#online_scores.append(t.score)
time_sync('tracker all')
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids))
#results.append((frame_id + 1, online_tlwhs, online_ids, online_scores))
# if show_image or save_dir is not None:
# online_im = vis.plot_tracking(img0, online_tlwhs, online_ids, frame_id=frame_id,
# fps=1. / timer.average_time)
# if show_image:
# cv2.imshow('online_im', online_im)
# if save_dir is not None:
# cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)), online_im)
frame_id += 1
if VIS_FLAG:
import matplotlib.pyplot as plt
fig = plt.figure(figsize=(16, 9))
ax = fig.add_subplot(221)
im1 = img0.copy()
rpn_map = torch.sigmoid(
out['rpn_map']).squeeze(0).squeeze(0).cpu().numpy()
rpn_map = cv2.resize(rpn_map, im1.shape[:2][::-1], cv2.INTER_CUBIC)
rpn_map = (rpn_map * 255).astype(np.uint8)
print(rpn_map.max(), rpn_map.min(), rpn_map.shape)
heat_map = cv2.applyColorMap(rpn_map, cv2.COLORMAP_JET)
im1 = (im1.astype(np.int16) + heat_map.astype(np.int16)) // 2
im1 = im1.astype(np.uint8)
im2 = cv2.cvtColor(im1, cv2.COLOR_BGR2RGB)
ax.imshow(im2)
ind = 0
ax = fig.add_subplot(223)
affs = stuffs['hm'][0][ind][0]
if affs.dim() == 3:
affs = affs.squeeze(1)
affs = affs.cpu().numpy()
# sim_map = cv2.resize(affs, ())
pos_inds = gt_rois[0] > 0
pos_ref_inds = gt_ref_rois[0] > 0
print(pos_inds.sum(), pos_ref_inds.sum())
sim_map = affs[pos_inds][:, pos_ref_inds]
import csv
fscon = lambda x: str(float(x))
fakemap = out['rpn_map'].flatten()[out['rois']
[0].flatten()].sigmoid()
with open(
'/home/toka/code/FairMOT/csv_files/%s_%06d.csv' %
(seq, frame_id), 'w') as fd:
rpmap = gt_ref_rois[1][pos_ref_inds]
pmap = gt_rois[1]
dw = csv.DictWriter(fd, ['id'] + list(map(str, rpmap)) +
['pos_sum', 'neg_sum', 'pred', 'ori'])
dw.writeheader()
dw.writerows([
dict([('id', str(iii))] +
[(str(k), fscon(v))
for k, v in zip(rpmap, affs[iin][pos_ref_inds])] +
[('pos_sum', fscon(affs[iin][pos_ref_inds].sum())),
('neg_sum',
fscon(affs[iin].sum() -
affs[iin][pos_ref_inds].sum())),
('pred', fscon(out['hm'][iin])),
('ori', fscon(fakemap[iin]))])
for iin, iii in enumerate(pmap)
])
sim_map = cv2.resize(sim_map, (1800, 600))
print(sim_map.shape, sim_map.max(axis=1))
ax.imshow(sim_map)
ax = fig.add_subplot(224)
affs = stuffs['hm'][0][ind][1]
if affs.dim() == 3:
affs = affs.squeeze(1)
affs = affs.cpu().numpy()
# sim_map = cv2.resize(affs, ())
sim_map = affs[pos_inds][:, pos_ref_inds]
sim_map = cv2.resize(sim_map, (1800, 600))
# print(sim_map.shape, sim_map.max(axis=1))
ax.imshow(sim_map)
fig.show()
plt.show()
if frame_id > 10:
break
# save results
import pickle
with open(result_filename + '.dets.pkl', 'wb') as fd:
pickle.dump(tracker.raw_dets, fd)
write_results(result_filename, results, data_type)
#write_results_score(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
bbox_dir=None,
show_image=True,
frame_rate=30):
if save_dir:
mkdir_if_missing(save_dir)
if bbox_dir:
mkdir_if_missing(bbox_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
for path, img0, detection in dataloader:
if frame_id % 1 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
online_targets, detection_boxes = tracker.update(detection, img0)
online_tlwhs = []
online_ids = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
#bbox detection plot
box_tlbrs = []
box_scores = []
img_bbox = img0.copy()
for box in detection_boxes:
tlbr = box.tlbr
tlwh = box.tlwh
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area:
box_tlbrs.append(tlbr)
box_scores.append(box.score)
timer.toc()
# save results
results.append(
([frame_id + 1] * len(online_tlwhs), online_tlwhs, online_ids))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
bbox_im = vis.plot_detections(img_bbox,
box_tlbrs,
scores=box_scores)
if show_image:
cv2.imshow('online_im', online_im)
cv2.imshow('bbox_im', bbox_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
cv2.imwrite(os.path.join(bbox_dir, '{:05d}.jpg'.format(frame_id)),
bbox_im)
frame_id += 1
# save results
track_pools = []
id_pools = []
write_results(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30,
conf_thres=0.3):
if save_dir:
mkdir_if_missing(save_dir)
tracker = GNNTracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
for i, (path, img, img0, p_img_path, p_img) in enumerate(dataloader):
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
if i == 0:
p_boxes, init_img_path, p_img = dataloader.initialize(
use_letter_box=opt.use_letter_box)
else:
p_boxes, p_img = prepare_prev_img(dataloader, online_targets, opt,
p_img)
if opt.use_roi_align:
p_crops = p_boxes.clone()
_, h, w = p_img.shape
p_crops = p_crops.cuda()
p_crops_lengths = [len(p_crops)]
edge_index = create_inference_time_graph(opt, p_boxes, p_crops,
p_img)
# convert boxes from xyxy to normalized according to p_img dimensions
p_crops[:, 0] = p_crops[:, 0] / w
p_crops[:, 1] = p_crops[:, 1] / h
p_crops[:, 2] = p_crops[:, 2] / w
p_crops[:, 3] = p_crops[:, 3] / h
online_targets = tracker.update(
blob,
img0,
p_crops,
p_crops_lengths,
edge_index,
gnn_output_layer=opt.inference_gnn_output_layer,
p_imgs=p_img.unsqueeze(0).cuda(),
conf_thres=conf_thres)
else:
p_crops = torchvision.ops.roi_align(input=p_img.unsqueeze(0),
boxes=[p_boxes],
output_size=opt.crop_size)
p_crops = p_crops.cuda()
p_crops_lengths = [len(p_crops)]
edge_index = create_inference_time_graph(opt, p_boxes, p_crops,
p_img)
online_targets = tracker.update(
blob,
img0,
p_crops,
p_crops_lengths,
edge_index,
gnn_output_layer=opt.inference_gnn_output_layer,
p_imgs=None,
conf_thres=conf_thres)
online_tlwhs = []
online_ids = []
online_confs = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
t_conf = t.score
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
online_confs.append(t_conf)
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids, online_confs))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
scores=online_confs,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
write_results(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename=None,
save_dir=None,
show_image=False,
frame_rate=10,
use_cuda=True):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
person_mot = PersonMOT()
results = []
frame_id = 0
rflag = False
# for path, img, img0 in dataloader:
for i, (path, img, img0) in enumerate(dataloader):
if i % 12 != 0:
continue
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
if use_cuda:
blob = torch.from_numpy(img).cuda().unsqueeze(0)
else:
blob = torch.from_numpy(img).unsqueeze(0)
online_targets = tracker.update(blob, img0)
online_tlwhs = []
online_ids = []
#online_scores = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
# online_scores.append(t.score)
timer.toc()
# 发送atc帧信息
rflag = atcs.send_info(opt.device_id, opt.url, online_ids,
online_tlwhs, opt.input_stream)
if rflag and i % opt.check_setp == 0:
person_mot.handle_crime(opt.device_id, opt.url, opt.trigger,
online_ids, online_tlwhs, img0)
# 是否渲染图像
if show_image or save_dir:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
results.append((frame_id + 1, online_tlwhs, online_ids))
# 得分
# results.append((frame_id + 1, online_tlwhs, online_ids, online_scores))
# 展示图像
if show_image:
cv2.imshow('online_im', online_im)
cv2.waitKey(1)
# 存储图像
if save_dir:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
if opt.save:
write_results(result_filename, results, data_type)
#write_results_score(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30,
use_cuda=True):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
frame_id = 0
#for path, img, img0 in dataloader:
out_queue = Deque(maxlen=6)
for i, (path, img, img0) in enumerate(dataloader):
#if i % 8 != 0:
#continue
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
time_clean()
time_sync('tracker all')
if use_cuda:
blob = torch.from_numpy(img).cuda().unsqueeze(0)
else:
blob = torch.from_numpy(img).unsqueeze(0)
with torch.no_grad():
time_sync('forward_half and queue')
out = tracker.model.backend.forward_half(
blob, roi_num=tracker.model.backend.roi_top_k)
out_ref = {}
out_ref.update(out)
out_ref['rois'], _ = tracker.model.backend.get_rois(
out['rpn_map'], 300)
refs = tracker.model.backend.forward_rois(out_ref)
out_queue.append(refs)
while len(out_queue) < out_queue.maxlen:
out_queue.append(refs)
time_sync('forward_half and queue')
time_sync('forward_all')
output, stuffs, _ = tracker.model.forward_half(out, out_queue)
time_sync('forward_all')
# info_debug(output)
# jj = output['rois'][0] % out['rpn_map'].shape[-1]
# ii = output['rois'][0] // out['rpn_map'].shape[-1]
# u = output['rpn_map'][:, :, ii, jj]
# print(output['hm'][:20])
# input()
# import pickle
# print(u.shape)
# print(u.flatten().sigmoid())
# print((u.flatten().sigmoid() - output['hm'].flatten().sigmoid()).abs().max())
# with open('two_none.pkll', 'wb') as fd:
# pickle.dump(output, fd)
# input()
time_sync('tracker update')
online_targets = tracker.update(blob, img0, output)
time_sync('tracker update')
online_tlwhs = []
online_ids = []
#online_scores = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
#online_scores.append(t.score)
time_sync('tracker all')
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids))
#results.append((frame_id + 1, online_tlwhs, online_ids, online_scores))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
import pickle
with open(result_filename + '.dets.pkl', 'wb') as fd:
pickle.dump(tracker.raw_dets, fd)
write_results(result_filename, results, data_type)
#write_results_score(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
polygon,
paths,
data_type,
result_filename,
frame_dir=None,
save_dir=None,
bbox_dir=None,
show_image=True,
frame_rate=30,
polygon2=None,
line1=None,
line2=None,
cam_id=None):
count = 0
if save_dir:
mkdir_if_missing(save_dir)
if bbox_dir:
mkdir_if_missing(bbox_dir)
if frame_dir:
mkdir_if_missing(frame_dir)
if cam_id is not None:
if cam_id is not None:
f = open('/data/submission_output/cam_' + str(cam_id) + ".txt",
"w")
else:
f = None
tracker = JDETracker(opt,
polygon,
paths,
frame_rate=frame_rate,
polygon2=polygon2)
timer = Timer()
results = []
frame_id = 1
for path, img, img0 in dataloader:
img0_clone = copy.copy(img0)
# if frame_id % 1 == 0:
# logger.info('Processing frame {} ({:.2f} fps)'.format(frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(
0) if opt.gpus[0] >= 0 else torch.from_numpy(img).cpu().unsqueeze(
0)
online_targets, detection_boxes, out_of_polygon_tracklet = tracker.update(
blob, img0)
if f is not None:
for frame_ind, _, track_type, mov_id in out_of_polygon_tracklet:
if mov_id != 'undetermine' and track_type != 'undetermine':
if track_type in [
'person', 'motor', 'motorcycle', 'bicycle',
"tricycle"
]:
track_type = 1
f.write('cam_' + str(cam_id) + ',' + str(frame_ind) +
',' + str(mov_id) + ',' + str(track_type) +
'\n')
elif track_type in ['car', 'van']:
track_type = 2
f.write('cam_' + str(cam_id) + ',' + str(frame_ind) +
',' + str(mov_id) + ',' + str(track_type) +
'\n')
elif track_type in ['bus']:
track_type = 3
f.write('cam_' + str(cam_id) + ',' + str(frame_ind) +
',' + str(mov_id) + ',' + str(track_type) +
'\n')
elif track_type in ['truck']:
track_type = 4
f.write('cam_' + str(cam_id) + ',' + str(frame_ind) +
',' + str(mov_id) + ',' + str(track_type) +
'\n')
online_tlwhs = []
online_ids = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
online_tlwhs.append(tlwh)
online_ids.append(tid)
#bbox detection plot
box_tlbrs = []
box_scores = []
box_occlusions = []
types = []
img_bbox = img0.copy()
for box in detection_boxes:
tlbr = box.tlbr
tlwh = box.tlwh
box_tlbrs.append(tlbr)
box_scores.append(box.score)
box_occlusions.append('occ' if box.occlusion_status ==
True else 'non_occ')
types.append(box.infer_type())
timer.toc()
# save results
for track in out_of_polygon_tracklet:
frame_idx, id, classes, movement = track
results.append((opt.input_video.split('/')[-1][:-4], frame_idx,
classes, movement))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time,
out_track=out_of_polygon_tracklet)
bbox_im = vis.plot_detections(img_bbox,
box_tlbrs,
scores=box_scores,
box_occlusion=None,
types=types)
if show_image:
cv2.polylines(online_im, [np.asarray(polygon)], True,
(0, 255, 255))
cv2.polylines(bbox_im, [np.asarray(polygon)], True, (0, 255, 255))
cv2.polylines(img0_clone, [np.asarray(polygon)], True,
(0, 255, 255))
cv2.imshow('online_im', online_im)
cv2.imshow('bbox_im', bbox_im)
# if save_dir is not None:
# cv2.polylines(online_im,[np.asarray(polygon)],True,(0,255,255))
# cv2.polylines(bbox_im,[np.asarray(polygon)],True,(0,255,255))
# # cv2.polylines(bbox_im,[np.asarray(paths['3'])],True,(0,255,255))
# # cv2.polylines(bbox_im,[np.asarray(paths['4'])],True,(0,255,255))
# if polygon2 is not None:
# cv2.polylines(online_im,[np.asarray(polygon2)],True,(0,0,255))
# cv2.polylines(bbox_im,[np.asarray(polygon2)],True,(0,0,255))
# if line1 is not None and line2 is not None:
# cv2.polylines(online_im,[np.asarray(line1)],True,(134,128,255))
# cv2.polylines(online_im,[np.asarray(line2)],True,(134,128,255))
# #cv2.polylines(img0_clone,[np.asarray(polygon)],True,(0,255,255))
# cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)), online_im)
# cv2.imwrite(os.path.join(bbox_dir, '{:05d}.jpg'.format(frame_id)), bbox_im)
# cv2.imwrite(os.path.join(frame_dir, '{:05d}.jpg'.format(frame_id)),img0_clone)
frame_id += 1
# if frame_id==150:
# BaseTrack._count=0
# return frame_id, timer.average_time, timer.calls
# save results
return frame_id, timer.average_time, timer.calls
def eval_seq(opt,
dataloader,
data_type,
result_filename,
save_dir=None,
show_image=True,
frame_rate=30):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
len_all = len(dataloader)
start_frame = int(len_all / 2)
frame_id = int(len_all / 2)
for i, (path, img, img0) in enumerate(dataloader):
if i < start_frame:
continue
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
img_pil = Image.open(path).convert('RGB')
normalize = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
img_norm = transforms.Compose(
[T.RandomResize([800], max_size=1333), normalize])
img_norm = img_norm(img_pil)
if frame_id % 20 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(
frame_id, 1. / max(1e-5, timer.average_time)))
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
online_targets = tracker.update(img_norm.cuda().unsqueeze(0), img0)
online_tlwhs = []
online_ids = []
#online_scores = []
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id
vertical = tlwh[2] / tlwh[3] > 1.6
if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
online_tlwhs.append(tlwh)
online_ids.append(tid)
#online_scores.append(t.score)
timer.toc()
# save results
results.append((frame_id + 1, online_tlwhs, online_ids))
#results.append((frame_id + 1, online_tlwhs, online_ids, online_scores))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0,
online_tlwhs,
online_ids,
frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)),
online_im)
frame_id += 1
# save results
write_results(result_filename, results, data_type)
#write_results_score(result_filename, results, data_type)
return frame_id, timer.average_time, timer.calls
def eval_seq(opt, dataloader, data_type, result_filename, gt_filename, save_dir=None, show_image=True, frame_rate=30):
if save_dir:
mkdir_if_missing(save_dir)
tracker = JDETracker(opt, frame_rate=frame_rate)
timer = Timer()
results = []
gts = []
frame_id = 0
tid_max, tid_temp = 1, 1
for path, img, img0 in dataloader:
if frame_id % 100 == 0:
logger.info('Processing frame {} ({:.2f} fps)'.format(frame_id, 1. / max(1e-5, timer.average_time)))
# if frame_id>400:
# break
if '0000001' in path:
tid_max = tid_temp
tracker = JDETracker(opt, frame_rate=frame_rate)
# run tracking
timer.tic()
blob = torch.from_numpy(img).cuda().unsqueeze(0)
# online_targets = tracker.update(blob, img0)
online_targets = tracker.update_sep3(blob, img0, conf_thres=[0.4, 0.3, 0.4, 0.4, 0.4, 0.4, 0.4, 0.5, 0.5]) ## use class-separated tracker
# print(online_targets)
online_tlwhs = []
online_ids = []
online_cids = [] #class id
for t in online_targets:
tlwh = t.tlwh
tid = t.track_id + tid_max
tcid = t.class_id
# vertical = tlwh[2] / tlwh[3] > 1.6
# if tlwh[2] * tlwh[3] > opt.min_box_area and not vertical:
if tlwh[2] * tlwh[3] > opt.min_box_area:
online_tlwhs.append(tlwh)
online_ids.append(tid)
online_cids.append(tcid)
tid_temp = max(tid, tid_temp)
timer.toc()
gt_tlwhs = []
gt_ids = []
gt_cids = []
gt_path = path.replace('images', 'labels_with_ids').replace('jpg', 'txt')
gt_targets = read_gt_txts(gt_path)
for (tlwh, tid, tcid) in gt_targets:
gt_tlwhs.append(tlwh)
gt_ids.append(tid)
gt_cids.append(tcid)
# save results
results.append((frame_id + 1, online_tlwhs, online_ids, online_cids))
# save gts
gts.append((frame_id + 1, gt_tlwhs, gt_ids, gt_cids))
if show_image or save_dir is not None:
online_im = vis.plot_tracking(img0, online_tlwhs, online_ids, frame_id=frame_id,
fps=1. / timer.average_time)
if show_image:
cv2.imshow('online_im', online_im)
if save_dir is not None:
cv2.imwrite(os.path.join(save_dir, '{:05d}.jpg'.format(frame_id)), online_im)
frame_id += 1
# save results
write_results(result_filename, results, data_type)
# save gts
write_results(gt_filename, gts, data_type)
return frame_id, timer.average_time, timer.calls
