A famous weakly supervised learning model for video anomaly detection (VAD).

original repository

• ShanghaiTech i3d features
• Revised few inconvenient codes
• i3d features
  • Extracted for every 16 frames
  • 10 crops augmentation means torchvision TenCrop

• 3d features extraction: Use x3d rather than i3d, c3d which are 3D conv models
  • Use torchvideo: pytorchvideo_x3d
  • UCF_Crime x3d feature extraction (requires more than 24 hrs)
  • Train RTFM with UCF_Crime + x3d dataset
    • Good performance!
    • Tested with new motion data (never seen): Got correct: white_check_mark:, Got correct: x:
      • normal:{drinking ✅, googling ❌, normal ✅, toilet ✅}
      • abnormal:{capture ❌, drawing ✅, writing ✅}
      • Things to notice: 1. Result changes even the data does not have big difference 2. Model works fairly good for webcam domain data even it is not trained with
      • If webcam domain data is given and weakly supervised, there is hope :)

• Webcam abnormal face data
• Webcam abnormal motion data

• RuntimeError: Expected a 'cuda' device type for generator but found 'cpu'
  • See cuda_troubleshooting.txt
• No visdom error
  • visdom is a web-based visualization tool
  • pip install visdom
  • (in a new terminal) python -m visdom.server
  • can see the result by accessing via web