Reproduce “Graph-Jigsaw Conditioned Diffusion Model for Skeleton-based Video Anomaly Detection” on OpenResearch

About this paper

Ali Karami, Thi Kieu Khanh Ho, Narges Armanfard

Addressing limitations in skeleton-based video anomaly detection, GiCiSAD from McGill University and Mila - Quebec AI Institute, integrates dynamic graph learning and a self-supervised graph-level jigsaw puzzle with a conditional diffusion model. This unified framework achieves state-of-the-art AUROC scores on four datasets, such as 78.0% on HR-STC and 89.6% on HR-Avenue, while reducing training parameters by 40% compared to prior leading unsupervised methods.

Key insights

• Dynamically learning graph structures for spatio-temporal dependencies provides a more sensitive representation for anomaly detection.
• Self-supervised graph-level jigsaw puzzles, particularly with inter-community shuffling, are effective for training models to discern subtle region-specific anomalies.
• Conditioning a diffusion model with dynamic graph representations enables the generation of diverse and context-aware future motion sequences, crucial for robust anomaly detection.

Summary by alphaXiv.