我发现了bug
Opened this issue · 0 comments
kongzijian commented
1.在.pkl的数据集中提到
data/ffpp_face_rects.pkl:
We use MTCNN detector to extract face bounding boxes for each frame for each real ("youtube") video and save it in the Python dict form and store it the pickle file ffpp_face_rects.pkl.这里面存储的是real video的信息,但是在video_dataset.py中采用的是所有数据。
2. 在video_dataset.py中采用帧反转的方法if random.random() < 0.5:frames = frames[::-1],但是未对sampled_idxs进行反转