Source-linked AI summary
WildMatch: Weakly Supervised Image Matcher Adaptation for Wildlife Re-Identification
Turhan Can Kargin, Piotr Kubaty, Ekaterina Rostovskaya, Izabela Wierzbowska, Bartosz Zieliński, Marcin Przewięźlikowski
TL;DR
Wildlife re-identification needs models to recognize distinctive local markings, but identity labels do not provide local correspondence annotations. WildMatch mines informative image pairs with a pretrained matcher and adapts it using identity labels; it improves retrieval across datasets and retains gains on unseen individuals.
Problem
Pretrained local matchers are not adapted to wildlife imagery, while identity labels do not identify corresponding local regions between images.
Source: §1 · p. 2Method
WildMatch mines high-scoring same-identity positives and different-identity hard negatives, then fine-tunes the matching module using identity supervision.
Source: §1 · p. 3Results
WildMatch improves over off-the-shelf matchers and a state-of-the-art local–global fusion approach across most datasets and candidate budgets, with gains persisting for unseen individuals.
Source: §1 · p. 3, §6 · p. 13Takeaways & Limitations
Identity annotations already available in monitoring datasets can support data-efficient specialization of local matchers to wildlife re-identification.
Source: §6 · p. 13
Abstract
from arXiv · showhide
Individual animal re-identification from camera-trap imagery is an instance retrieval problem central to non-invasive wildlife monitoring: a query image must retrieve the correct individual from a reference set of known animals. This requires computer vision models to recognize distinctive local patterns in fur, skin, or other visual markings. Current approaches either learn global embeddings as a classification problem, requiring many labeled images per individual while largely ignoring local evidence, or apply off-the-shelf, domain-agnostic image matchers. Although such matchers are pretrained on large and diverse image collections, adapting them to wildlife imagery is challenging because available datasets are small and lack correspondence-level annotations. We study weakly supervised adaptation of a pretrained keypoint matcher using only identity labels, without keypoint-level or geometric correspondence ground truth. We mine informative image pairs with the pretrained matcher, derive weak positive and negative supervision from identity agreement, and contrastively fine-tune the matching network to strengthen correspondences for same-identity pairs and suppress them for different identities. Across open-source wildlife re-identification datasets, our approach improves accuracy over off-the-shelf matchers and a state-of-the-art local--global fusion method. Under an open-world protocol with held-out individuals, it learns a transferable correspondence prior rather than memorizing training identities. To our knowledge, this is the first study of matcher-level, identity-supervised adaptation for animal re-identification. Our method enables data-efficient specialization of image matching models to wildlife domains using identity annotations already available in typical monitoring datasets.
1 Introduction
Wildlife re-identification depends on matching distinctive local markings, but identity labels do not specify which image regions correspond. WildMatch uses those labels to adapt pretrained matchers and improves retrieval across diverse datasets, including for unseen individuals.
1 Introduction
- Local feature matching is motivated by distinctive animal markings that global image embeddings may obscure through pooling.Human experts compare local coat patterns, while global embeddings discard much of this local evidence.Source: §1 · p. 2
- Identity labels establish whether images show the same animal but not which local structures correspond, and opposite flanks may share little visible evidence.The central question is whether a general-purpose matcher can be specialized using the sparse identity annotations already available in re-identification datasets.Source: §1 · p. 3
- WildMatch mines high-scoring same-identity positives and different-identity hard negatives, then fine-tunes the matcher using identity agreement rather than correspondence annotations.The pretrained matcher identifies plausible local evidence; identity labels determine whether to encourage or suppress it.Source: §1 · p. 3
- WildMatch improves over off-the-shelf local matchers and a state-of-the-art local–global fusion approach across most datasets and candidate budgets, with gains transferring to unseen individuals.The evaluation spans eight wildlife re-identification datasets; the open-world results indicate a transferable correspondence prior rather than a closed-set identity classifier.Source: §1 · p. 3
2 Related Work
Wildlife re-identification research includes global image representations and local matching of distinctive markings. WildMatch differs by adapting the matcher itself using animal identity labels.
2 Related Work
- General-purpose learned matching pipelines combine local feature extraction with neural modules that reason jointly over image pairs, while global re-identification methods learn image representations.The related work describes LightGlue, RDD, and LoMa among general-purpose matching methods, and species-specific or multi-species global models.Source: §2 · p. 4
- Wildlife systems include global representations, local matching, and fusion approaches, but the matcher in these approaches remains general-purpose rather than adapted with animal identity labels.Examples include HotSpotter, WildFusion, and AnimalCLEF systems using LightGlue-based local verification.Source: §2 · pp. 4–5
- Prior matcher adaptation fine-tunes sparse matchers for compatibility across feature extractors, whereas WildMatch adapts matching to wildlife imagery using image-level identity agreement.WildMatch does not require geometric, pose, or keypoint-level correspondence supervision.Source: §2 · p. 5
3 Method
WildMatch adapts pretrained local matchers using identity labels alone, mining informative pairs before fine-tuning the matching module. At inference, correspondence scores rerank retrieved gallery candidates.
3 Method
- WildMatch mines positive and hard-negative pairs with a pretrained matcher, then fine-tunes its matching module using an identity-based triplet ranking objective.The approach retains the original local feature representation and applies to pretrained matchers with a learnable matching module.Source: §3 · p. 5
3.1 Problem Setup and Inference
- At inference, the matcher scores each query–candidate pair from retained local correspondences, and candidates are reranked by decreasing score.Candidate images are drawn from a global-encoder shortlist; background masking may reduce uninformative keypoints and stationary-camera background bias.Source: §3.1 · pp. 5–6
3.2 Identity-Guided Pair Mining
- Pair mining selects high-scoring same-identity positives and different-identity hard negatives because same-animal images may show opposite flanks with little shared local evidence.Mining is performed once before fine-tuning on the training set; up to five high-scoring examples are selected from each pool, and anchors without an eligible positive are discarded.Source: §3.2 · p. 6
3.3 WildMatch: Identity-Supervised Matcher Adaptation
- Training samples positives and a mixture of hard and random negatives, updates only the matching module, and keeps the local feature extractor and global encoder frozen.A relaxed score computed before discrete correspondence selection supplies a training signal even when inference filters would retain no correspondences; the triplet margin objective favors higher scores for same-identity pairs.Source: §3.3 · pp. 6–7
4 Experimental Setup
WildMatch is evaluated on eight wildlife datasets using closed-set and unseen-identity retrieval protocols. Comparisons cover local matchers, global retrieval, local–global fusion, and identity classification.
4 Experimental Setup
- Eight public datasets span diverse species; database images form both the retrieval gallery and the only fine-tuning data, while query images are held out.For Salamander, queries are the latest capture dates for individuals photographed on at least two dates.Source: §4 · p. 7, Table 1 · p. 8
- The datasets retain noisy frames and are highly imbalanced, motivating evaluation that accounts for performance across rare and frequent individuals.Examples include overexposed, blurred, corrupted, animal-free, or marking-obscured images.Source: §4 · p. 7, Figure 3 · p. 8
- The unseen-identity CzechLynx protocol tests 44 individuals absent from training, using earliest encounters as galleries and later encounters as queries.It provides 160 gallery and 2,081 query images; k=160 is exhaustive matching.Source: §4 · p. 7
- Baselines include cosine retrieval with MegaDescriptor-L or DINOv3, default LoMa and RDD matchers, WildFusion, and MegaDescriptor-L identity classifiers.Top-5 accuracy is primary; balanced top-1 accuracy weights rare and frequent individuals equally.Source: §4 · p. 8
- WildMatch fine-tunes only the 11.9M-parameter matching module, with detector and descriptor frozen, for 300 epochs; CzechLynx training takes 5.1 GPU-hours.Inputs are 512-pixel images with up to 512 keypoints, and mined triplets are generated once using the pretrained matcher.Source: §4 · p. 9
5 Experimental Results
Across datasets, matcher adaptation improves retrieval over default checkpoints and WildFusion, while component, classifier, and cross-identity tests clarify its scope. Qualitative examples show correspondences linking distinctive markings across viewing conditions.
5.1 Retrieval Accuracy
- Across eight datasets, WildMatch improves over the default matcher and WildFusion at almost every candidate budget, and the default matcher needs two to five times longer candidate lists to reach its accuracy.Re-ranking with any matcher also lifts accuracy well above embedding retrieval alone.Source: §5.1 · p. 9
5.2 Which Component Should Be Adapted?
- Fine-tuning the matching module improves both matcher architectures, whereas descriptor fine-tuning lowers balanced top-1 accuracy; tuning both is slightly better but costs 16–21 times more.A trainable descriptor prevents feature caching and changes the matching module’s inputs at every step, so other experiments adapt only the matching module.Source: §5.2 · p. 10
5.3 Matcher Adaptation or Identity Classification
- On CzechLynx, the matcher reaches 35% top-5 accuracy after 5 GPU-hours, while the fully fine-tuned classifier takes 10 GPU-hours merely to reach the unadapted matcher’s top-5 accuracy.The classifier remains below 20% balanced top-1 accuracy and cannot recognize individuals outside its label set.Source: §5.3 · p. 10
5.4 Generalization to Unseen Identities
- On unseen CzechLynx identities, both fine-tuned matchers outperform their defaults and WildFusion at every budget in both metrics, including exhaustive k=160.The result supports a transferable species-specific matching prior rather than memorization of training identities; an identity classifier cannot be applied to individuals without a training class.Source: §5.4 · p. 11, Table 3 · p. 11
5.5 Applicability Beyond LoMa
- The adaptation also improves RDD-LightGlue over its default checkpoint, although the advantage narrows at the largest candidate budgets.The authors attribute the narrowing to training on mostly hard negatives, leaving easier candidates in longer lists less represented.Source: §5.5 · p. 12, Figure 6 · p. 12
5.6 Qualitative Results
- Correct top-1 retrievals show correspondences linking coat, skin, or shell markings across viewpoint and illumination changes, and those matches can be inspected by experts.Examples include hyena and leopard spots, nyala stripes, and salamander patches.Source: §5.6 · p. 12, Figure 7 · p. 13
6 Conclusion
WildMatch specializes local matchers to wildlife re-identification using identity labels available in monitoring datasets. Its gains extend across datasets and to unseen individuals, while exposing the correspondences behind retrievals.
6 Conclusion
- WildMatch specializes general-purpose local matchers for wildlife re-identification using identity labels without keypoint or geometric correspondence annotations.The approach adapts the matching module rather than requiring correspondence-level labels.Source: §6 · p. 13
- Across eight datasets, matching-module adaptation improves LoMa over its pretrained checkpoint and a competitive local–global fusion method at almost every candidate budget, and experiments show extension beyond one matcher architecture.The conclusion cites RDD as evidence that the approach is not limited to LoMa.Source: §6 · p. 13
- The adapted matcher retains gains on individuals unseen during training and exposes retrieval correspondences, supporting use with newly observed individuals without retraining.The paper describes this as a data-efficient specialization that needs no keypoint or geometric annotations.Source: §6 · p. 13