Reproduce “Sequential keypoint density estimator: an overlooked baseline of skeleton-based video anomaly detection” on OpenResearch

About this paper

Anja Delić, Matej Grcić, Siniša Šegvić

Researchers at the University of Zagreb developed SeeKer, a skeleton-based video anomaly detection method that models human motion by autoregressively factorizing skeleton sequence density at the keypoint level, integrating keypoint detection uncertainty into its interpretable anomaly scores. SeeKer achieved an AUROC of 77.9% and AP of 80.3% on the UBnormal dataset, surpassing prior skeleton-based approaches by over 6 percentage points in AUROC and 17 percentage points in AP.

Key insights

• Autoregressive factorization of skeleton sequence density at the keypoint level effectively captures the causal and compositional nature of human motion, outperforming skeleton-level and monolithic approaches.
• Incorporating keypoint detection confidence scores into the anomaly scoring mechanism is essential for robustness and significantly improves detection performance.
• Learning the covariance matrix for the multivariate normal distributions of keypoints is critical for accurate modeling, leading to over 10% absolute AUROC improvement compared to fixed covariance.

Summary by alphaXiv.