Abstract
Weakly supervised video anomaly detection (WS-VAD) aims to train models using only video-level labels while simultaneously achieving precise localization of anomalous segments. Existing methods universally assume that input videos are captured under favorable imaging conditions. However, in real-world deployments, adverse illumination conditions such as foggy illumination, rain, and nighttime cause severe image degradation, leading to significant feature distribution shift and substantially degrading detection performance. Illumination-robust weakly supervised video anomaly detection built upon the Information Bottleneck (IB) architecture. Our framework introduces three complementary innovations: (1) an Illumination-Conditioned Normalization layer (WCN) that dynamically adjusts illumination-adaptive affine parameters through a lightweight illumination-aware branch, effectively counteracting the contamination of feature distributions caused by illumination degradation; (2) a Motion-Aware Temporal Enhancement module (MATE) that extracts motion cues via frame differencing and injects them into the feature sequence via gated modulation, alleviating the over-compression of motion-based anomaly signals by the information bottleneck; and (3) a Temporal Reasoning Module (TRM) that re-models temporal dependencies through a lightweight cross-frame attention mechanism, restoring the model’s capability for precise temporal localization of anomalous behaviors. Comprehensive experiments on ShanghaiTech Campus, UCF-Crime, and XD-Violence benchmarks demonstrate that RobustNet significantly outperforms the DS-IB Net baseline and prior state-of-the-art methods under clear, foggy, rainy, nighttime, and mixed illumination conditions, while ablation studies verify the effectiveness and complementarity of all three innovations. RobustNet is also efficient: the three modules add only