HanwooReID: Multi-view cattle re-identification with pose-aware transformer enhancements

Identifying individual Hanwoo cattle automatically is particularly challenging: Hanwoo cattle exhibit highly similar visual characteristics, and unlike pedestrian re-identification, cattle undergo dramatic changes in appearance and scale as their poses and camera viewpoints vary. To address these challenges, we introduce HanwooReID, a large-scale multi-view dataset collected on real farms, comprising 9,929 images of 31 cattle captured under diverse viewpoints, poses, and lighting conditions. Building on this dataset, we propose a transformer-based framework that integrates a pose-guided heatmap encoder (PHE) to focus attention on identity-relevant regions and a viewpoint-constrained retrieval (VCR) strategy that projects hoof keypoints onto a bird's-eye view (BEV) plane to estimate cattle orientation. This BEV-based orientation estimation effectively filters out gallery candidates with inconsistent viewpoints, substantially improving matching accuracy under severe pose and view variations. Extensive experiments on both closed-set and open-set protocols show that our method outperforms existing baselines, achieving up to 7.4% and 6.8% improvements in mean Average Precision (mAP), respectively, demonstrating its effectiveness for precision livestock farming.