Multi-Camera Video Anomaly Detection Dataset Based on PETS 2009 Benchmark Dataset

This repository contains a preprocessed multi-camera dataset for multi-camera video anomaly detection task. The dataset consists of 4 overlapped camera video scenes.

Dataset Description

The dataset is provided in .npy format (numpy version: 1.24.3). Each .npy file contains a list of video scenes for a single camera video.

Dataset Contents

The repository contains the following files:

• Training splits for each camera:

  • pets2009-comb-train-view-001.npy
  • pets2009-comb-train-view-002.npy
  • pets2009-comb-train-view-003.npy
  • pets2009-comb-train-view-004.npy

• Test Splits for each camera:

  • pets2009-comb-test-view-001.npy
  • pets2009-comb-test-view-002.npy
  • pets2009-comb-test-view-003.npy
  • pets2009-comb-test-view-004.npy

• DATA-LOADING.ipynb: A jupyter notebook providing a demonstration on how to load and process the dataset.

Dataset Structure

The data in each .npy file is organized as a list of dictionary objects, where each object represents a video scene. Each object contains the following properties:

• name: A string representing the name of the video scene;
• X_i: Each video of each camera corresponds to a matrix with dimension $n$ x $2048$, where $n$ is a variable number of existing clips and the number of attributes is $1024$ - $1024$ Inflated 3D (i3D) deep features for appearance - RGB for each clip in the video);
• y_i: the category of video scene ($0.0$ or $1.0$);
• y_fi: A matrix $n$ x $16$ representing the frame labels.

Citation

If you use this dataset in your research or project, please cite it as:

@misc{pereira2023mil,
      title={A MIL Approach for Anomaly Detection in Surveillance Videos from Multiple Camera Views}, 
      author={Silas Santiago Lopes Pereira and José Everardo Bessa Maia},
      year={2023},
      eprint={2307.00562},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

ArXiv URL: https://arxiv.org/abs/2307.00562