SG-NSA: A Training-Free Inference Enhancement Framework for Dead-Cattle Re-Identification

In complex ranch environments, ReID of cattle faces remains highly challenging due to postmortem appearance degradation, severe occlusion, environmental interference, and the extremely limited number of samples available for each individual. These factors substantially undermine the reliability and robustness of existing visual ReID models. To address these challenges, we propose SG-NSA (Semantic-guided NFC Set Aggregation), a training-free inference-stage enhancement framework built upon a trained and fixed ReID backbone. Without introducing additional learnable parameters, modifying the backbone architecture, or fine-tuning the trained model, SG-NSA improves dead-cattle face ReID performance through feature-level correction, set-level aggregation, and semantic-guided filtering. SG-NSA consists of three synergistic modules: Neighborhood Feature Correction (NFC), which alleviates feature drift; Image Set Aggregation (ISA), which constructs stable set-level identity representations; and Semantic-guided Filtering (SF), which constrains and progressively narrows the candidate identity space. Together, these modules form a progressively enhanced inference mechanism. Furthermore, we construct a real-world live-to-dead paired cattle-face dataset collected over two years from ranches in Inner Mongolia, comprising 19,809 images of 1385 individual cattle. Experimental results demonstrate that SG-NSA consistently yields substantial gains across multiple baseline models. When applied to TransReID-SSL, it achieves 81.6% mAP and 80.7% Rank-1 accuracy, corresponding to improvements of 65.2% and 64.8%, respectively, over the baseline.