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

Maayan Yesharim, R. G. Bina Perl+4 more|arXiv (Cornell University)|Jan 13, 2026

Amphibian and Reptile Biology0 citations

TL;DR

The paper shows zero-shot deep local-feature matching yields near-perfect photo-identification of the Hula painted frog, outperforming global-feature models and enabling a fast, two-stage retrieval workflow for open-set identification.

ABSTRACT

Accurate individual identification is essential for monitoring rare amphibians, yet invasive marking is often unsuitable for critically endangered species. We evaluate state-of-the-art computer-vision methods for photographic re-identification of the Hula painted frog (Latonia nigriventer) using 1,233 ventral images from 191 individuals collected during 2013-2020 capture-recapture surveys. We compare deep local-feature matching in a zero-shot setting with deep global-feature embedding models. The local-feature pipeline achieves 98% top-1 closed-set identification accuracy, outperforming all global-feature models; fine-tuning improves the best global-feature model to 60% top-1 (91% top-10) but remains below local matching. To combine scalability with accuracy, we implement a two-stage workflow in which a fine-tuned global-feature model retrieves a short candidate list that is re-ranked by local-feature matching, reducing end-to-end runtime from 6.5-7.8 hours to ~38 minutes while maintaining ~96% top-1 closed-set accuracy on the labeled dataset. Separation of match scores between same- and different-individual pairs supports thresholding for open-set identification, enabling practical handling of novel individuals. We deploy this pipeline as a web application for routine field use, providing rapid, standardized, non-invasive identification to support conservation monitoring and capture-recapture analyses. Overall, in this species, zero-shot deep local-feature matching outperformed global-feature embedding and provides a strong default for photo-identification.

Motivation & Objective

• Motivate non-invasive individual identification for critically endangered amphibians.
• Evaluate state-of-the-art computer-vision methods for photo-ID on the Hula painted frog.
• Compare zero-shot deep local-feature matching with deep global-feature embeddings.
• Develop a scalable two-stage workflow that balances accuracy and computational efficiency.
• Demonstrate practical deployment through a web application for field use.

Proposed method

• Assess 1,233 ventral images from 191 individuals (2013–2020 capture-recapture) with zero-shot deep local-feature matching.
• Compare local-feature matching to deep global-feature embedding models.
• Fine-tune global-feature models to establish a stronger baseline (top-1 and top-10 metrics).
• Implement a two-stage workflow: global-feature retrieval plus re-ranking with local features.
• Quantify end-to-end runtime reduction from several hours to minutes while preserving high accuracy (~96% top-1 on labeled data).
• Provide a web-application deployment for routine field use.

Experimental results

Research questions

• RQ1 Can zero-shot deep local-feature matching achieve higher photo-ID accuracy than global-feature embeddings for the Hula painted frog?
• RQ2 What is the impact of fine-tuning global-feature models on top-1 and top-10 identification performance?
• RQ3 Does a two-stage retrieval pipeline (global followed by local features) maintain high accuracy while reducing runtime?
• RQ4 Is the system capable of open-set identification through separation of match scores for same- versus different-individual pairs?
• RQ5 Can the pipeline be effectively deployed as a practical, non-invasive field tool?

Key findings

• Zero-shot deep local-feature matching achieves 98% top-1 closed-set identification accuracy.
• Fine-tuned global-feature model reaches 60% top-1 (91% top-10) but remains below local-feature performance.
• Two-stage workflow reduces end-to-end runtime from 6.5–7.8 hours to ~38 minutes while maintaining ~96% top-1 closed-set accuracy.
• Thresholding on same- vs different-individual score distributions enables open-set identification.
• Web application deployment supports rapid, standardized, non-invasive identification for conservation monitoring.