with torch.no_grad():
for video_name in bridge.video_index.keys():
video_index = bridge.video_index[video_name]
for video_name in bridge.video_index.keys():
frame_num_list = bridge.video_index[video_name]
ids = [
bridge.video_frame_2_id(video_name, frame_num)
for frame_num in frame_num_list
]
video_dets = torch.Tensor()
# 将每一个视频片段转换为tensor
for frame_num in frame_num_list:
frame_dets = bridge.get_frame_det(video_name, frame_num)
for det in frame_dets:
det_t = torch.tensor([
det["image_id"], det["bbox"][0], det["bbox"][1],
det["bbox"][2], det["bbox"][3], det["score"]
]).unsqueeze(dim=0)
if video_dets.shape[0] == 0:
video_dets = det_t
else:
video_dets = torch.cat([video_dets, det_t])
# 选取这段视频中的score最大的框
used_mask = torch.zeros(video_dets.shape[0])
max_score, max_score_index = select_max_det(video_dets, used_mask)
used_mask[
max_score_index] = True # 标记已经被选择的框，排除后迭代，直到max score小于阈值0.2
