Pose-Aware Proxies for Unsupervised Marine Wildlife Re-Identification

ICLR 2026 Conference Submission20736 Authors

Keywords: unsupervised re-identification, pose-aware proxies, contrastive learning, wildlife monitoring, thresher shark dataset

TL;DR: We curate the first thresher shark dataset and propose pose-aware proxies for unsupervised re-ID in ecological video, improving discrimination while exposing cross-pose retrieval as the key challenge.

Abstract: Scaling wildlife re-identification remains challenging due to the need for expert-defined anatomical landmarks (manual photo-ID) and large labeled datasets (supervised learning). In Malapascua, Philippines, which is home to endangered thresher sharks, expertise and funding are limited, yet divers capture abundant unlabeled footage. To make this resource usable, we curate the first structured dataset of thresher shark dive videos, organized via co-occurrence and track-based local identities. Leveraging this corpus, we introduce pose-aware proxies: coarse orientation labels that provide weak viewpoint supervision within a clustering-based contrastive framework. We evaluate without global identity labels using three field-aligned metrics: within-track consistency (WTC), co-occurrence recall (CoR@k), and mutual-exclusion error (MEError@k). On our dataset, the TP6 variant improves temporal stability and reduces impostor matches, while slightly lowering CoR at small k. These results show that pose-conditioned guidance extends proxy-based unsupervised learning to unconstrained ecological video, prioritizing precision over immediate recall, and they isolate cross-pose matching as a key open challenge for future work.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 20736