Weakly Supervised Anomaly Detection with RGB and Flow Features

This work is inspired by the paper "Batch Normalized Weakly Supervised Anomaly Detection". However, unlike the original, which uses only RGB features, this implementation incorporates both RGB and Optical Flow features to improve training stability.

Dataset and Features

• XD-Violence
  We achieved an 84.92% score on the XD-Violence dataset, which is comparable to using RGB-only features. However, the inclusion of Flow features leads to a more stable training process.

  • 📥 Download the XD-Violence I3D features from this link

• UCF-Crime
  Note: UCF-Crime does not provide Flow features by default.

  • 📥 Download the UCF-Crime I3D RGB features from this link

  To extract Flow features for UCF-Crime:

  • Either download a pretrained I3D Flow model
  • Or use the provided script in feature_extract/ to extract Flow features from your video files.

Feature Extraction

The feature_extract/ directory contains a script to extract Optical Flow features from videos.

Make sure to organize your input videos in a folder, then run:

python video2flow2i3d.py --src_dir ../../UCF_actual/Anomaly-Videos/Burglary/ --output_dir ./your_directory/Burglary/

This script will compute the optical flow and extract I3D features for each video.

Training and Testing

To train and test the model using the extracted features, run:

python main.py --root_dir ./XD_violence/i3d-features/