vp = Path(args.video_path)
assert vp.is_file(), '{} not a file'.format(vp)
cap = cv2.VideoCapture(str(vp))
assert cap.isOpened(), 'Cannot open video file!'
cv2.namedWindow('yolo', cv2.WINDOW_NORMAL)
while True:
# Decode
ret, frame = cap.read()
if not ret:
break
# Inference
tic = time.perf_counter()
dets = yolo.detect_get_box_in([frame], box_format='ltrb')
toc = time.perf_counter()
print('infer duration: {:0.3f}s'.format(toc - tic))
dets = dets[0]
# Drawing
show_frame = frame.copy()
for det in dets:
ltrb, conf, clsname = det
l, t, r, b = ltrb
cv2.rectangle(show_frame, (int(l), int(t)), (int(r), int(b)),
(255, 255, 0))
cv2.putText(show_frame,
'{}:{:0.2f}'.format(clsname, conf), (l, b),
cv2.FONT_HERSHEY_DUPLEX,
fontScale=1,
od = YOLO( bgr=True, score=od_thresh, batch_size=bs,
model_path='model_data/pp_reanchored_best_val.h5',
anchors_path='model_data/PP_ALL_anchors.txt',
classes_path='model_data/PP_classes.txt' )
remainder = len(pp)
print('Forward passes...')
for start_idx in tqdm(range(0, len(pp), bs)):
if remainder < bs:
bs = remainder
od.regenerate(bs)
images, impaths = pp.pull_images(start_idx, bs)
remainder -= len(impaths)
results = od.detect_get_box_in(images, box_format='ltrb')
det_dict = {}
for i, result in enumerate(results):
impath = impaths[i]
for det in result:
box_info, score, pred_cls = det
pred_cls_id = od.class_names.index(pred_cls) + 1
l,t,r,b = box_info
this_array = np.array([l,t,r,b,score])
if pred_cls_id not in det_dict:
det_dict[pred_cls_id] = {}
if impath not in det_dict[pred_cls_id]:
det_dict[pred_cls_id][impath] = this_array
else:
det_dict[pred_cls_id][impath] = np.vstack([det_dict[pred_cls_id][impath], this_array])
# Note that all images of a batch needs to be same size for the post-processing to make sense.
for impath in ip.rglob('*'):
if impath.suffix in ['.png', '.jpg', '.jpeg']:
img = cv2.imread(str(impath))
if first_size is None:
first_size = img.shape[:2]
else:
img = cv2.resize(img, first_size[::-1])
images.append(img)
image_paths.append(impath)
else:
raise Exception('Path given does not exist')
# Inference
tic = time.perf_counter()
all_dets = yolo.detect_get_box_in(images, box_format='ltrb')
toc = time.perf_counter()
print('infer duration: {:0.3f}s'.format(toc - tic))
# Drawing
for img, imgpath, dets in zip(images, image_paths, all_dets):
show_img = img.copy()
for det in dets:
ltrb, conf, clsname = det
l, t, r, b = ltrb
cv2.rectangle(show_img, (int(l), int(t)), (int(r), int(b)),
(255, 255, 0))
cv2.putText(show_img,
'{}:{:0.2f}'.format(clsname, conf), (l, b),
cv2.FONT_HERSHEY_DUPLEX,
fontScale=1,
