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