This repository is the official PyTorch implementation of "Integrating Non-local Memory-Augmented Reconstruction and Appearance-Motion Frame Prediction for Video Anomaly Detection"

python==3.6
pytorch==1.5.1
mmcv-full==1.3.1
mmdet==2.11.0
scikit-learn==0.23.2
edflow==0.4.0
PyYAML==5.4.1
tensorboardX==2.4

Please follow the read.me to prepare the dataset.

First, to train the NL_ML_MemAE , run:

$ python nl_ml_memAE_train.py

Then, to train the AMTAE with reconstructed flows, run:

$ python trian.py

Finally, finetune the whole framework as:

$ python finetune.py

To evaluation the trained model, run:

$ python eval.py --model_save_path=[model_save_path] 
                 --cfg_file=[cfg_file] 

You can download the pretrained weights of NMR-AFP for Ped2, Avenue and ShanghaiTech datasets from here.

This code is heavily borrowed from hf2vad. Thanks LiUzHiAn for their contributions.