Sequential Spatial Graphs for Video Anomaly Detection

Big Data Research Project - CentraleSupélec 2022/23

Authors: Niccolò Morabito and Yi Wu

Data

The following three video anomaly detection benchmarks have been used:

• Avenue
• ShanghaiTech
• StreetScene

For each of them, we built:

• Yolo-annotated txt files in data/yolo_annotated_datasets/ (by using the code available in this repo);
• training set pickle files which only contains normal videos in data/training_graphs/;
• video parameters pickle files with the video information (like width and height of the frame) in data/video_parameters/;
• testset pickle files which contains both normal and abnormal videos in data/testing_graphs/;
• testset labels pickle files in data/testing_labels/.

For the complete data/ folder, please check the following GoogleDrive link.

Code

The project is vided into the following folders:

• src/grah_generation/ for the generation of NetworkX graphs starting from Yolo-annotated datasets files;
• src/anomaly_generation/ for the corruption of graphs;
• src/embedding_training/ for GCN and transformers;
• src/common_utils/ for other useful code.

In order to train the model, it is sufficient to run the code contained in theh src/pipeline.ipynb Jupyter notebook.

A demo is also available in the src/demo.ipynb file to show the process from the YOLO-annoted dataset to a prediction.