A Dual-Stream CLIP–ViT Framework for Open-Set Animal Re-Identification: Multi-Seed Ablation, Background-Bias Bracketing, and Query-Time Robustness Analysis

Animal re-identification (Re-ID) asks whether two images show the same individual, a recognition task that fits naturally into applications such as reuniting lost pets with their owners. Existing methods report strong scores, but typically under a single seed, one mask granularity, and no query-time corruption analysis, leaving open whether the gains survive deployment. We propose a hierarchical framework decoupling localisation (a YOLOv8 soft-crop) from identity embedding: a dual-stream network fusing a frozen CLIP ViT-B/16 (learned projection) with a fine-tuned ViT-Base carrying L2-norm part attention, trained under ArcFace. On a combined cat+dog open-set benchmark of 173 identities, it attains Rank-1 0.9742/mAP 0.8597 over three seeds, surpassing a ViT-only ablation by +2.39 Rank-1 and +2.07 mAP. Open-set verification shows all configurations converge near 68% true acceptance at the strictest false-acceptance rate. A background-bias evaluation brackets the embedding’s background reliance between a bounding-box lower bound and a SAM-silhouette upper bound; a manual audit retains the reliable soft crop. A nine-corruption audit identifies down-sampling and motion blur as dominant. On PetFace, the architecture retrieves across 14,716 unseen identities and remains viable in a few-shot regime. A Descriptor Vector Exchange (DVE) extension is Pareto-dominated, traced to the ViT’s coarse feature map and architectural redundancy.