Reproduce “CHAD: Charlotte Anomaly Dataset” on OpenResearch

About this paper

Armin Danesh Pazho, Ghazal Alinezhad Noghre, Babak Rahimi Ardabili, Christopher Neff, Hamed Tabkhi

The Charlotte Anomaly Dataset (CHAD) is introduced as a large-scale, high-resolution, multi-camera dataset for video anomaly detection in a commercial parking lot setting. It provides over 1.15 million frames, including over one million normal frames and comprehensive person-level annotations such as bounding boxes, identity tracking, and human keypoints, to standardize input for skeleton-based methods. Benchmarking shows state-of-the-art models achieve average AUC-ROC scores up to 0.718 on CHAD, with the dataset presenting a higher challenge compared to existing benchmarks.

Key insights

• CHAD's extensive normal behavior data and pre-generated person annotations directly address critical needs for robust unsupervised and consistent skeleton-based anomaly detection research.
• The dataset's
• extremely feature rich
• videos and inclusion of subtle, complex anomalies make it significantly more challenging for current models than existing benchmarks.
• Employing a multi-metric evaluation approach (AUC-ROC, AUC-PR, EER) offers a more holistic understanding of anomaly detection model performance, particularly beneficial for imbalanced datasets.

Summary by alphaXiv.