HIERARCHICAL PROGRESSIVE REFINEMENT FOR WEAKLY SUPERVISED VIDEO ANOMALY DETECTION

Abstract

Weakly Supervised Video Anomaly Detection (WSVAD) is a technique that uses coarse video-level labels to find transient anomalous actions. Recent architectures are more likely to use heavy spatial-temporal graph networks or large Vision-Language Models (VLMs) to give semantic reasoning and address inherent pseudo-label noise. Nevertheless, constructing joint-embedding spaces directly between text and video modalities is computationally costly and often results in overfitting. More to the point, these methods are trying to learn the visual signature of anomalies, which is a false assumption to make since abnormal events are unlimited and heterogeneous in nature; normalcy, though, is circumscribed and extremely repetitive. To address this, we present a hierarchical progressive refinement framework for weakly supervised video anomaly detection. In the first stage, we present a lightweight Text-Conditioned Multiple Instance Learning (TC-MIL) model. Instead of enforcing a more complicated contrastive align ment, TC-MIL deploys semantic language as a literal gating mechanism. We obtain motion-sensitive visual representations with a pre-trained I3D backbone and produce textual representations with a frozen CLIP encoder. The projection and multiplica tion of the text embedding of the temporal sequence results in the mathematical per formance of the language as a semantic filter enhancing the use of visual channels of interest to the desired anomaly and actively suppressing non-harmful background noise. In the second stage, we present CORE-MIL (Compact Orthogonal Residual Ex traction), a simple, mathematically based framework that reverses this paradigm. In place of attempting to determine chaos, CORE-MIL explicitly measures the inability to rebuild order. We inculcate a small, learnable Normalcy Anchor Matrix to learn common background behaviors. This subspace receives the incoming video features and the reconstructed normal component is subtracted from the raw input. This gives an orthogonal (residual) which is flat to daily activities but steep in case of unforeseen events. Extensive testing on the UCF-Crime and XD-Violence datasets demonstrates this framework delivers highly competitive, interpretable localization accuracy. It reduces computational costs by aggressively screening background noise before classification and avoiding cross-modal alignment, doing so while slashing computational overhead and training time by an order of magnitude.