A Benchmark and Cross-Camera Protocol for Open-Set Re-Identification of Holstein Cattle: Synchronized Multi-View Capture and Viewpoint-Invariant Alignment

Public cattle re-identification benchmarks remain limited in scale and rarely provide leakage-controlled open-set and cross-camera evaluation. Using 25,209 curated crops of 197 Holstein cows recorded by four synchronized side-view cameras in a passageway of a working farm, we investigate viewpoint-related shifts in the embedding space and evaluate cattle re-identification under camera-disjoint gallery/query construction. The extractor combines an animal-pretrained MegaDescriptor Swin-B backbone with multi-level projections, learned gating, quality weighting, and two-level aggregation. Under the strictly inductive identity-disjoint cross-camera protocol without camera centering, the proposed model achieved Rank-1 of 55.7% (95% CI: 47.7–63.8%), Rank-5 of 90.0% (86.3–93.7%), and mAP of 70.2% (66.3–74.1%). Under the same protocol, TransReID ViT-B/16 achieved 38.6%, 77.9%, and 55.9%, while OSNet-x1.0 achieved 12.9%, 47.9%, and 30.7%, respectively. The comparison with the pure-GAP configuration was not statistically resolved (Holm-adjusted p = 0.91). In held-out open-set evaluation, the validation-selected global cosine threshold obtained the strongest mean AUROC, F1, and AUOSCR (63.2%, 66.9%, and 44.4%); the corresponding results were 60.2%, 64.4%, and 38.8% for Mahalanobis rejection and 56.9%, 63.2%, and 41.7% for EVT. Zero-shot evaluation at a second corridor on the same farm yielded Rank-1 of 26.6%, Rank-5 of 41.9%, and mAP of 35.3%. These results establish cross-camera feasibility under the defined protocols and same-farm cross-installation transfer, but do not establish cross-farm, cross-device, seasonal, longitudinal, or full production robustness.