Paper & Code¶
WildMatch: Weakly Supervised Image Matcher Adaptation for Wildlife Re-Identification. Turhan Can Kargin, Piotr Kubaty, Ekaterina Rostovskaya, Izabela Wierzbowska, Bartosz Zieliński, Marcin Przewięźlikowski. Preprint, arXiv:2610.07384, 2026.
The preprint is released under CC BY 4.0.
Citation¶
@misc{kargin2026wildmatch,
title={WildMatch: Weakly Supervised Image Matcher Adaptation for Wildlife Re-Identification},
author={Turhan Can Kargin and Piotr Kubaty and Ekaterina Rostovskaya and Izabela Wierzbowska and Bartosz Zieliński and Marcin Przewięźlikowski},
year={2026},
eprint={2610.07384},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.07384}
}
Acknowledgements¶
This research was supported by the flagship project "Artificial Intelligence Computing Center Core Facility" of the Priority Research Area DigiWorld under the Strategic Programme Excellence Initiative at the Jagiellonian University. We gratefully acknowledge Poland's high-performance computing infrastructure PLGrid (HPC Center: ACK Cyfronet AGH) for providing computer facilities and support within computational grant no. PLG/2026/019483.
We thank the creators of the datasets used in this work: the CzechLynx team, the authors of WildlifeReID-10k and of the datasets it brings together, the organisers of AnimalCLEF2025, who allowed us to show some SalamanderID2025 photos, and the authors of the Kaggle Jaguar Re-ID competition, who allowed us to use its data for research. Background masks were computed with SAM 3 (Meta).
Contact¶
Questions about the method or the code: open an issue in the evaluation repository linked below.
Code¶
One repository implements the whole pipeline: pair mining, matcher fine-tuning and evaluation.
| Repository | Role | Link |
|---|---|---|
WildMatch |
Pair mining, matcher fine-tuning (LoMa, RDD-LightGlue), evaluation: probes, candidate lists, metrics, run manifests, paper tables and figures, this page | GitHub |
Repository map¶
src/wildmatch/: the installable package behind thewildmatchcommand.src/wildmatch/evaluate/: probe dispatch (cosine, WildFusion, Vismatch matchers, linear and efficient probes), metrics (top-k, balanced top-1, mAP) and stable ranking.src/wildmatch/matchers/: Vismatch matcher profiles and custom-checkpoint loading.src/wildmatch/mining/: pair mining (wildmatch mine): dataset views, feature caches, strong-match indices.src/wildmatch/matcher_finetune/: matcher fine-tuning (wildmatch finetune-matcher): triplet sampling, relaxed score, training recipe, checkpoints.src/wildmatch/data/: dataset views, COCO-RLE masking, split safety checks and dataset preparation (wildmatch prepare).src/wildmatch/sweep/: evaluation grids as immutable Slurm or local submissions.src/wildmatch/reporting/: run manifests, run index, paper tables and figures.src/wildmatch/conf/: Hydra configuration, path profiles and the dataset registry.paper/: page exporters and paper figures.tests/: dependency-light regression tests.
Third-party components¶
| Component | Use | Pinning |
|---|---|---|
| MegaDescriptor-L | Global embeddings for the candidate list and classifier baselines | wildlife-tools release |
| DINOv3-L | Cosine-retrieval baseline | facebook/dinov3-vitl16-pretrain-lvd1689m (gated) |
| LoMa | Matcher | via Vismatch, pinned commit (evaluation); lomatch, pinned commit (mining, fine-tuning) |
| RDD, LightGlue | Matcher | via Vismatch, pinned commit (evaluation); RDD code vendored unchanged (mining, fine-tuning) |
| WildFusion | Baseline | wildlife-tools |
| SAM 3 | Background removal for SalamanderID2025 | prompt "Salamander" |
The code is Apache-2.0 and the fine-tuned matcher checkpoints are CC BY-NC 4.0. MegaDescriptor-L, used for every method's candidate list, is CC BY-NC 4.0, so the full pipeline is for non-commercial use. Each dataset keeps its own license (see Datasets).