Readme.md

LAVAD

This method uses frame data as input. You need to firstly generate a annotations/ file and a test.txt under it, which should be like:

0000_ball_anomaly_free 0 240 0
0001_ball_anomaly_free 0 240 0
0002_ball_anomaly_free 0 240 0
0003_ball_leak 0 240 0
...

where the first column is the name of the video file, and the third is the total frame number of the video. The second and the fourth are just 0. The annotations file should at the same level of the frames file, that is:

frame_data/
├─ ball/
│  ├─ training/
│  │  ├─ frames/
│  │  ├─ annotations/
│  │  │  ├─ test.txt

After running the script you will find some new files like captions, index etc. generated. They won't influence the original frames file.

MemAE

This method uses frames as input. You should firstly generate a frames_idx folder by running the src/MemAE/matlab_script/matlabrunner.py, the frames_idx folder should be at the same level as the frames folder.

MGFN

This method uses i3d feature. Two lists of train or test video feature paths are required. Take test list for object ball for example, it should be like this:

path_to_your_data/ball/i3d/test/leak0000.npy
path_to_your_data/ball/i3d/test/leak0001.npy
path_to_your_data/ball/i3d/test/leak0002.npy
path_to_your_data/ball/i3d/test/leak0003.npy
path_to_your_data/ball/i3d/test/leak0005.npy
...

The final list should be a list file like test_ball.list.

VadCLIP

This method uses clip feature. This method also need two csvs like MGFN, but we have automated this step. You only need to change the feature root path in src/VadCLIP/process.py. If you want to modified to your own setting (e.g. the ratio of normal and abnormal instances) you can also look into src/VadCLIP/process.py to change them.

S3R

This method uses i3d feature. The preparation for this method is relatively complex. You can find the preparation step in Data preparation, i3d feature part.