Scalable animal re-identification without labels is essential for wildlife monitoring, especially in resource-limited settings; however, most unsupervised re-ID frameworks remain limited to human datasets. This study systematically evaluates three state-of-the-art frameworks—Self-paced Contrastive Learning (SpCL), Cluster Contrast (CC), and Transformer-Based Multi-Granular Features (TMGF)—on the NDD20 dolphin dataset, a curated underwater image collection featuring white-beaked dolphins (Lagenorhynchus albirostris).

Dolphin viewpoints were manually annotated to address the lack of camera ID labels and camera-aware proxies were substituted with pose-aware proxies to isolate pose variation. All frameworks were trained fully unsupervised using clustering-derived pseudo-labels and contrastive objectives, with ground-truth identities reserved solely for evaluation. Retrieval performance was assessed using mean Average Precision (mAP) and Cumulative Matching Characteristic (CMC) scores. TMGF consistently outperformed SpCL and CC, boosting mAP by 3% and demonstrating greater robustness to pose variation and intra-class variability. View-specific evaluation outperformed aggregated retrieval, suggesting that flank-dependent identity cues are significant for dolphin re-ID. In contrast, SpCL and CC, despite competitive Top-10 accuracy, exhibited lower mAP, indicating reduced consistency.

This study offers the first comprehensive assessment of unsupervised re-ID models on a marine wildlife dataset. It reveals that pose-aware proxies are effective for species with view-invariant or bilaterally consistent identifiers (e.g., humans, dorsal fins, tail flukes), but less so for species with asymmetric or view-dependent cues (e.g., flank markings). These findings underscore the importance of species-aware design when adapting unsupervised learning to ecological domains, advancing the development of AI-driven tools for biodiversity monitoring and marine conservation.