Illumination-Adaptive Multi-Branch Feature Fusion for Robust Person Re-Identification

Abstract: Person re-identification (Re-ID) across camera networks remains sensitive to illumination variation, since a single global appearance encoding is rarely robust to both bright and poorly lit capture conditions. This paper presents an adaptive illumination-robust Re-ID pipeline that estimates a continuous per-crop illumination score, generates four parallel enhanced views of each detected person (raw, CLAHE, gamma-corrected, and multi-scale Retinex), extracts an OSNet embedding for each view, and fuses the four embeddings using illumination-conditioned weights before performing cosine-similarity gallery retrieval via FAISS. The system is evaluated on Market-1501 under a stratified protocol comparing a Normal split against a synthetically generated Lowlight split, and comparing the proposed adaptive fusion against a static/average-fusion ablation baseline. Results show Rank-1 accuracy saturated at 98-99.6% across all four conditions, while mean Average Precision (mAP) is more discriminative, falling from approximately 71% on the Normal split to 44-46% on the Lowlight split. Contrary to the design hypothesis, the heuristic adaptive fusion does not outperform static average fusion in this evaluation, and static fusion is marginally better under Lowlight conditions. These findings are presented transparently as evidence that the current rule-based fusion weighting requires learning-based replacement, motivating a concrete direction for future work rather than an unqualified robustness claim.

Keywords: person re-identification, illumination robustness, adaptive feature fusion, OSNet, YOLOv8, Retinex, CLAHE, FAISS, Market-1501