0. Install SlowFast (https://github.com/facebookresearch/SlowFast).

1. Prepare the pre-trained model.

   1. Download the SlowFast Network pre-trained on Kinetics-400 dataset.
      See 1-prepare_model/model_urls.txt. Put SLOWFAST_8x8_R50.pkl in pretrainings/.
   2. Put slowfast_installation_dir/configs/Kinetics/SLOWFAST_8x8_R50.yaml in pretrainings/.
   3. Convert '.pkl' format checkpoint to '.pth' file.
      Cd 1-prepare_model/ and run cvt_pkl2pth.sh. Then SLOWFAST_8x8_R50.pth is generated in pretrainings/.

2. Prepare datasets.

   1. Download Avenue, ShanghaiTech and Corridor datasets.
      See 2-prepare_data/dataset_urls.txt.
   2. Cd 2-prepare_data/ and run extract_frames.py to extract frames for videos. Use '--help' to see the usage.
      ST: only need to do this for the training videos.
      Avenue: extract frames for both training & testing videos.
      Corridor: use mvfile_corridor.py to put all the training/testing videos in one directory first, and then extract frames (use --skip_first option to skip the first black frame).
   3. Convert the frame-level labels to '.npz' file.
      We have done this cumbersome step. The '.npz' files can be seen in groundtruths/.
      '.npz' file: keys=video names; values=arrays, 0 for normality, 1 for anomaly.

3. Extract features.

   1. ShanghaiTech & Avenue
      Cd 3-extract_features/for_ST_Avenue/. Modify and run run.sh to extract features for ST and Avenue datasets.
   2. Corridor
      Cd 3-extract_features/for_Corridor/.
      Run recrop.py first to extract 3 crops of each frame. Use '--help' to see the usage.
      Then run run.sh to extract features for Corridor dataset.
   3. Generate snippet-level-packaged features for all the datasets.
      Step 1) and step 2) generated video-level-packaged features. However, snippet-level-packaged features will also be used afterwards.
      For all the training & testing datasets, cd 3-extract_features/ and run snippet_level_packaged.py to convert video-level-packaged features to snippet-level-packaged features.

Now, for each dataset, we have got 1) video-level-packaged features, 2) snippet-level-packaged features, and 3) frame-level lables. We will use these features and lables in the training and testing codes.