Vision-based Re-identification (ReID) is crucial for intelligent surveillance yet remains vulnerable to adverse-weather degradations such as rain and fog, which simultaneously corrupt visual clarity and identity-specific cues. Existing image-enhancement and robust-feature paradigms struggle when multiple degradations co-occur, while recent CLIP-based ReID models have scarcely examined cross-modal alignment under weather distortions. To bridge this gap, we propose ScA-UniReID, a Scene-Aware Degradation Universal ReID framework built upon CLIP’s dual-encoder architecture. ScA-UniReID introduces dual textual prompts—target-oriented for identity features and degradation-oriented for weather noise—and an adaptive control module that dynamically re-weights them to disentangle identity semantics from degradation artifacts. Extensive experiments on pedestrian and maritime ReID benchmarks under diverse adverse-weather protocols show that ScA-UniReID outperforms state-of-the-art methods and generalizes robustly to unseen conditions, validating its efficacy and universality.
Wei et al. (Fri,) studied this question.