感谢作者的工作！
有一个对于temporal_triplet_loader.py脚本中训练的细节想请教：
在生成训练用数据集中，调用的self.sample_normal_abnormal_clips()方法中，即./untils/dataloaders/init.py line97-line110中。

                for start in range(0, length - inv):
                    end = start + inv
                    video_clips = images_paths[start:end:t]
                    reversed_video_clips = list(reversed(video_clips))

                    # check is normal or abnormal
                    if len(frame_mask) != 0 and np.count_nonzero(frame_mask[start:end:t]) >= at_least_anomalies:
                        video_abnormal_clips.append(video_clips)
                        video_abnormal_clips.append(reversed_video_clips)
                    else:
                        video_normal_clips.append(video_clips)
                        video_normal_clips.append(reversed_video_clips)

                print('sample video {} at time {}.'.format(v_name, t))

使用reversed_video_clips = list(reversed(video_clips))生成反向的视频段，是否是单纯的为了扩充数据，还是有其他的考虑与细节呢？
再次感谢您的工作！
祝好！

Thanks for the great work!
When I check the path "MLEP/data/annotations/", in the "shanghaitech_semantic_annotation.json" file, it seems to mark all the abnormal events in scene 1.
Could you provide the json annotation files of other 11 scenes?
thx again!

Why can the frame level annotation reproduction of my shanghaitech dataset only be about 70%? Can you provide your valuable advice

Feel your great work! thank you

In the code, you add parts of the test data to the train data. The new train data is "train_val_clips_dict".

In the "train_val_clips_dict", we have "normal" and "abnormal".

In the "abnormal", each frame is belong to abnormal.

And I have a question about the "normal".

check is normal or abnormal

if len(frame_mask) != 0 and np.count_nonzero(frame_mask[start : end : t]) >= at_least_anomalies:
video_abnormal_clips.append(video_clips)
video_abnormal_clips.append(reversed_video_clips)
else:
video_normal_clips.append(video_clips)
video_normal_clips.append(reversed_video_clips)

Some "video_clips" (five time steps) with one or two abnormal sample also belong to the "normal". This make me doubt.

Why we make these "video_clips" belong to "normal".

It seems that the link for download is not available now.
Could you please check it for onedrive users?

Which model is the backend that was proposed in the paper? I assume it would be cyclegan_convlstm, however from what I understand the proposed MLEP architecture does not use a cycle GAN and so I find the name confusing.

Is cyclegan_convlstm the proposed model @StevenLiuWen ?

Also, does training with only normal data correspond to the training scheme from Future Pred and Future Pred*?

Where is margin learning embodied in the paper? I found features in the code that you didn't take advantage of. I want to know how to use margin learning here?

Thx a lot!
I had downloaded all the pre-train model which were provided by link. But without the model with temporal annotation on ShanghaiTech. It was not uploaded. Could you pls upload the pre-train model with temporal annotation on SHTech?
Thx again！