CaTFormer: Prior-Guided Transformer with Hierarchical Supervision for Baggage Re-Identification

Baggage re-identification (ReID) aims to retrieve the same baggage instance across non-overlapping cameras or viewpoints, providing a practical tool for intelligent baggage management in structured transportation surveillance scenarios. Compared with pedestrian and vehicle ReID, baggage ReID is more challenging because baggage items often exhibit homogeneous appearances, limited semantic structures, and subtle identity cues that are sensitive to viewpoint, illumination, blur, and background changes. To address these challenges, we propose CaTFormer, a prior-guided transformer framework for baggage ReID. CaTFormer incorporates camera-aware and textual-attribute embeddings into Swin-based visual representation learning. An Adaptive Multi-Embedding Fusion (AMEF) module coordinates camera, color, and type priors before injecting the fused prior representation into visual tokens. In addition, a Candidate-Selected Semantic-Dominant Hierarchical Loss (CS-HL) supervises multi-level Swin features using a fixed stage-weight vector selected from a constrained candidate set. Experiments on the self-collected BaggageID dataset show that CaTFormer achieves 62.0% mAP, 85.4% Rank-1 accuracy, and 94.0% Rank-5 accuracy, outperforming the compared open-source ReID models under the same protocol. On the public MVB benchmark, CaTFormer obtains 87.0% mAP and 87.6% Rank-1 accuracy. Additional evaluations provide supplementary evidence for the applicability of CaTFormer and its CS-HL-based reduced variant in broader ReID settings.