Jigsaw-VAD

Official pytorch implementation for the paper entitled "Video Anomaly Detection by Solving Decoupled Spatio-Temporal Jigsaw Puzzles" (ECCV 2022)

Environments

• python 3.7.10
• pytorch 1.7.1
• torchvision 0.8.2
• scipy 1.7.1
• opencv-python 4.5.4.58
• pillow 8.2.0

Data Preparation

Please make sure that you have sufficient storage.

python gen_patches.py --dataset shanghaitech --phase test --filter_ratio 0.8 --sample_num 9

Training

python main.py --dataset shanghaitech --val_step 100 --print_interval 20 --batch_size 192 --sample_num 9 --epochs 100 --static_threshold 0.2

Testing

python main.py --dataset shanghaitech/avenue/ped --sample_num 9/7/7 --checkpoint xxx.pth

We provide the pre-trained weights for STC, Avenue and Ped2.

Citation

@inproceedings{wang2022jigsaw-vad,
  title     = {Video Anomaly Detection by Solving Decoupled Spatio-Temporal Jigsaw Puzzles},
  author    = {Guodong Wang and Yunhong Wang and Jie Qin and Dongming Zhang and Xiuguo Bao and Di Huang},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2022}
}