RELIABILITY-AWARE PROTOTYPE CALIBRATION FOR FROZEN POSE-FLOW VIDEO ANOMALY DETECTION

In one-class video anomaly detection, pose-flow detectors often rely on a single likelihood score to rank skeleton windows. However, these scores can be sensitive to pose-observation noise and may fail to capture the multimodal nature of normal human behavior. This paper addresses the challenge of improving anomaly detection performance when the detector backbone, cached skeleton tracks, and evaluation pipeline are already fixed and cannot be retrained.

The authors introduce Reliability-Aware Prototype Calibration (RPC), a post-hoc, training-free method to refine anomaly rankings. RPC operates by summarizing normal training data into a set of prototypes within the frozen latent space. For any test window, it calculates a 'prototype deviation'—the distance to the nearest normal prototype—and adds this to the standardized likelihood score. To handle the inherent noise in skeleton data (e.g., occlusions or tracking errors), RPC uses a reliability gate. This gate, derived from the keypoint confidence scores provided by the upstream pose estimator, modulates the contribution of the prototype deviation term. This ensures that unreliable pose data does not disproportionately influence the final anomaly score.

RPC was evaluated across two frozen pose-flow backbones and four datasets (ShanghaiTech, UBnormal, and their human-related variants). The method consistently improved frame-level AUROC in all eight tested scenarios, with gains ranging from 0.34 to 4.49 percentage points and an average improvement of 2.03 points. Ablation studies confirmed that the prototype deviation serves as the primary corrective signal for identifying abnormal behavior, while the reliability gating mechanism is particularly effective in scenarios where pose observations are less trustworthy.

This work provides a lightweight, effective solution for practitioners who need to improve existing surveillance systems without the high computational cost of retraining deep neural networks or re-processing raw video data. By treating anomaly detection as a score-level calibration problem, the authors demonstrate that significant performance gains can be achieved by better utilizing the information already present in a frozen model's latent space.