def get_sample(self, idx):
"""
preprocess one sample from the list
"""
cfg = self.cfg
img_file_path = self.imgs[idx]
imgname = os.path.splitext(os.path.basename(img_file_path))[0]
frame = cv2.imread(img_file_path)
im = frame.astype("float32")
resized_image = resizeImage(im, cfg.short_edge_size, cfg.max_size)
scale = (resized_image.shape[0] * 1.0 / im.shape[0] + \
resized_image.shape[1] * 1.0 / im.shape[1]) / 2.0
return resized_image, scale, imgname, (im.shape[0], im.shape[1])
def _run(self):
cfg = self.cfg
frame_count = 0
while frame_count < self.num_frame:
if self.stop_signal.is_set():
return
if self.is_moviepy:
suc = True
frame = next(self.vcap)
else:
suc, frame = self.vcap.read()
if not suc:
frame_count += 1
continue
if frame_count % self.frame_gap != 0:
frame_count += 1
continue
# process the frames
if self.is_moviepy:
# moviepy ask ffmpeg to get rgb24
frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
im = frame.astype("float32")
resized_image = resizeImage(im, cfg.short_edge_size, cfg.max_size)
scale = (resized_image.shape[0] * 1.0 / im.shape[0] + \
resized_image.shape[1] * 1.0 / im.shape[1]) / 2.0
self.queue.put((resized_image, scale, frame_count), block=True)
frame_count += 1
self._wait_queue()
if self.stop_signal.is_set():
# We're done
return
# skip some frame if frame_gap >1
if cur_frame % args.frame_gap != 0:
cur_frame+=1
continue
# 4. run detection on the frame stack if there is enough
im = frame.astype("float32")
if args.test_frame_extraction:
frame_file = os.path.join(video_out_path, "%d.jpg"%cur_frame)
cv2.imwrite(frame_file, im)
cur_frame+=1
continue
resized_image = resizeImage(im, args.short_edge_size, args.max_size)
scale = (resized_image.shape[0]*1.0/im.shape[0] + resized_image.shape[1]*1.0/im.shape[1])/2.0
feed_dict = model.get_feed_dict_forward(resized_image)
if args.get_box_feat:
sess_input = [model.final_boxes, model.final_labels, model.final_probs, model.fpn_box_feat]
final_boxes, final_labels, final_probs, box_feats = sess.run(sess_input,feed_dict=feed_dict)
assert len(box_feats) == len(final_boxes)
# save the box feature first
featfile = os.path.join(feat_out_path, "%d.npy"%(cur_frame))
np.save(featfile, box_feats)
else: