C2ST for Video Anomaly Detection

This repo contains the example code for BMVC2018 paper Classifier Two-Sample Test for Video Anomaly Detections.

It contains some implementation based on instructions in this paper since its code is not available. If you find our paper of this implementation is useful to your research, please cite

@inproceedings{liu2018classifier, title={Classifier Two-Sample Test for Video Anomaly Detections}, author={Yusha Liu and Chun-Liang Li and Barnab{\'a}s P{\'o}czos}, booktitle={BMVC}, year={2018} }

Using the code

• Download package:

$ git clone https://github.com/MYusha/Video-Anomaly-Detection

• Assume the default path is Video-Anomaly-Detection/pipeline.

Requirements: The code is written in Matlab 2017a, and used with laptop with MacOS. Please first install liblinear matlab (files included). And download pretrained vgg model to put inside /PrepareData/Appearance_feature/ for appearance feature extraction.

Specification: The functions for motion features computation are modified from re-implementation in https://github.com/gongruya/abnormality-detection.

Dataset preparation

Please put the Avenue datatset from CUHK inside the /Avenue_Dataset folder. Note that as mentioned we exclude the two videos which contains only abnormal events, since that contradicts with our assumption.

Instuctions

Generate scores: The experiment and parameters are included in the /pipeline/Run_script.m. Running this script will generate a series of features and anomaly score files for the videos.

Compute AUC: The script evaluation.m will read in the generated score files and compare with ground truth provided, to compute and display the AUC score. Individual AUC scores are also avaliable but not displayed.