Near-perfect photo-ID of the Hula painted frog with zero-shot deep local-feature matching

Accurate individual identification is essential to monitor rare species. Yet, invasive marking is often unsuitable for critically endangered species. Here, we evaluated state-of-the-art computer-vision methods for photographic re-identification of the Hula painted frog ( Latonia nigriventer ), a re-discovered critically endangered amphibian. Using 1232 ventral images from 191 individuals collected during 2013–2020 capture–recapture surveys, we compared deep local-feature matching in a zero-shot setting with deep global-feature embedding models. The local-feature matching pipeline achieved 99.8% top-1 closed-set identification accuracy, outperforming all global-feature models. Fine-tuning improved the best global-feature model to 62.1% top-1 (73.8% top-3) but remained far below local-matching performance. To combine scalability with accuracy, we implemented a two-stage workflow: a fine-tuned global-feature model retrieves a short candidate list that is re-ranked by local-feature matching. This reduced end-to-end runtime on our dataset with a single GPU from 6.3 h to 40 min while maintaining 99.0% top-1 closed-set accuracy. Separation of match scores between same- and different-individual pairs supports thresholding for open-set identification, enabling practical handling of novel individuals. We verified our approach with a prospective forward-in-time open-set evaluation on an unlabeled dataset (2016–2025) in which each query is matched only against earlier-dated references; at the calibrated operating threshold, 92 of 93 seen individuals were correctly matched to their prior identity while reducing expert review workload by 65%. We deploy this pipeline as a web application for routine re-identification of the Hula painted frog in the field, providing rapid, standardized, non-invasive identification to support conservation monitoring and capture–recapture analyses. Overall, zero-shot deep local-feature matching outperformed global-feature embedding and provides a strong default for photo-identification in this species.