Adaptive Temporal Modeling for Weakly Supervised Video Anomaly Detection

A new framework addresses variable anomaly durations in video surveillance by replacing rigid temporal priors with a Temporal Refinement Module. This module utilizes dynamic positional encoding and learnable class tokens to model long-range dependencies and improve snippet-level prediction stability.