Video anomaly detection in unconstrained surveillance environments remains challenging due to three intertwined issues: (i) modality imbalance among RGB, optical flow, and audio streams, (ii) cross-source asynchrony that disrupts segment-level correspondence, and (iii) weak supervision where only video-level labels are provided, making reliable evidence attribution difficult. To tackle these problems in a unified manner, we introduce HCC-VAD ( H ierarchical C ontrastive C alibration for V ideo A nomaly D etection), we first employ a progressive convolutional interaction aggregator that hierarchically compresses and mixes multimodal features, balancing information flow while reducing redundancy and training instability. These three levels are complementary: temporal consistency stabilizes each stream locally, neighborhood agreement preserves cross-modal semantic coherence under loose synchronization, and normal/abnormal separation improves the global discriminability of anomaly evidence. On top of this, a lightweight high-order interaction fusion head models multiplicative tri-modal co-occurrence patterns, amplifying audio–visual anomalies such as collisions, fights, and explosions. The additional calibration and fusion components are lightweight in practice because they operate on compact clip-level embeddings rather than dense raw features, introducing only modest training and inference overhead.
Zhou et al. (Fri,) studied this question.