WildMatch checkpoints
Fine-tuned matcher checkpoints for wildlife individual re-identification (WildMatch): LoMa-B and RDD-LightGlue, matching module only, epoch 299, one pair per dataset.
Paper: WildMatch: Weakly Supervised Image Matcher Adaptation for Wildlife Re-Identification (arXiv:2610.07384). Project page: wildmatch.gmum.net.
Layout: <dataset>/<matcher>/model.safetensors, plus czechlynx_protocol.json training
provenance next to the CzechLynx weights. SHA256SUMS.md lists every file's SHA-256 and set.
Two sets of RDD-LightGlue checkpoints
- Default (what
wildmatchuses since October 2026): for seven datasets (Hyena, Leopard, Sea star, Whale shark, Turtle, Salamander, CzechLynx open), RDD-LightGlue was retrained with the same recipe as LoMa (relaxed pair score, AdamW, LoMa-mined pairs, epoch 299). These files are in<dataset>/rdd-lightglue-shared-recipe/. For the other datasets (Nyala, CzechLynx closed) and for every LoMa checkpoint, the default is the paper's file. - Paper: the RDD-LightGlue checkpoints behind the paper's RDD-LightGlue results, unchanged in
<dataset>/rdd-lightglue/. Use these to reproduce the paper's numbers.
The retrained checkpoints are not in the paper; the project page compares them with the paper's (Beyond the paper).
Download and verify with the wildmatch package:
wildmatch weights download --dataset salamander # one dataset, or omit --dataset for all
wildmatch weights download --set paper # the paper's RDD-LightGlue checkpoints
wildmatch weights verify --set all
Licence
These checkpoints are released under the Creative Commons Attribution-NonCommercial 4.0 International licence (CC BY-NC 4.0): you may share and adapt them for non-commercial purposes, with attribution.
Why non-commercial:
- Training data: they were fine-tuned on data with non-commercial terms (WildlifeReID-10k).
- Base weights: the LoMa-B base weights were trained on non-commercial datasets.
- Pipeline: the evaluation pipeline uses MegaDescriptor-L (CC BY-NC 4.0) for its candidate lists.
The RDD-LightGlue base weights are Apache-2.0. The code that trains and evaluates the checkpoints is Apache-2.0. The datasets keep their own terms; the images are not included here.
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}
}