wjtan99 / bert_anomaly_video_classification

Overlooked Video Classification in Video Anomaly Detection

Overlooked Video Classification in Video Anomaly Detection

The power of video-level classification in almost all previous video anomaly detection is either overlooked or not studied well explicitly. With addition of a BERT or LSTM video classification, we achieve new SOTA results on UCF-Crime, ShanghaiTech, and XD-Violence datasts.

MIL-BERT

Standard MIL with BERT video classification

RTFM-BERT

Modificed RTFM with BERT video classification

bert_anomaly_video_classification's Issues

Regarding 'dataset.py'

When referring to the UCF-crime dataset within 'dataset.py', two txt files are called:

train_file = 'train_normal.txt' test_file = 'test_normalv3.txt'

I cannot find these txt files in any of the UCF-dataset versions or txt file splits.

Where can I find these files?

Thanks!

some questions about extracting I3D features

Will you extract I3D features of 10crop? If you have code, can you provide me with a copy? Thank you very much!

Code to generate the XDViolence_gt.npy needed for test_10crop.py

Could you please elaborate on how to create the XDViolence_gt.npy file to replicate your results using the XD-Violence dataset?

About the video_list generate

Thanks for sharing the code of this excellent work. But I have some questions about the code. It will be much appreciated if you could resolve my puzzles.

the result of your process of the ucf dataset and generating the video_list.txt in "/BERT_Anomaly_Video_Classification/tree/main/MIL-BERT/fix_ucf_crime_test_list.ipynb" is below:

['Abuse/Abuse028_x264.mp4|1412|[165, 240, -1, -1]', 'Abuse/Abuse030_x264.mp4|1544|[1275, 1360, -1, -1]', 'Arrest/Arrest001_x264.mp4|2374|[1185, 1485, -1, -1]', 'Arrest/Arrest007_x264.mp4|3144|[1530, 2160, -1, -1]', 'Arrest/Arrest024_x264.mp4|3629|[1005, 3105, -1, -1]']

but the result was processed by me is below：
Abuse/Abuse028_x264.mp4|1413|[165, 240, -1, -1]
Abuse/Abuse030_x264.mp4|1545|[1275, 1360, -1, -1]
Arrest/Arrest001_x264.mp4|2375|[1185, 1485, -1, -1]
Arrest/Arrest007_x264.mp4|3145|[1530, 2160, -1, -1]

there are different in num_frames, I use the "frames.pkl" released by https://github.com/junha-kim/Learning-to-Adapt-to-Unseen-Abnormal-Activities

Thanks very much