acceptodds
Under review as a conference paper at ICLR 2027

MetaFauna: A Multi-Species Animal Re-Identification Dataset with Structured Metadata

Abstract

Animal re-identification (ReID) aims to recognize the same individual across images, supporting non-invasive wildlife monitoring and ecological studies. In camera-trap imagery, however, identity-relevant visual evidence varies with observation conditions: viewpoint and occlusion affect which cues are visible, while illumination, blur, and scale affect how clearly those cues are captured. Most existing animal ReID datasets provide only images and identity labels, requiring models to distinguish identity-related differences from variation associated with these observation conditions using visual input alone. Recent work has begun incorporating auxiliary metadata, such as temperature, capture time, and orientation, into animal ReID. However, most of these attributes describe the temporal or environmental context of capture rather than the observation conditions of each image. To address this gap, we construct MetaFauna, a multi-species animal ReID dataset comprising 44,317 images of 2,500 individuals across eight species. Each image is paired with structured metadata describing its observation conditions, grouped into visual evidence visibility and capture conditions, and, where available, its environmental context. To study whether explicit descriptions of observation conditions improve animal ReID, we further introduce MetaRouter, a reference framework that conditions visual feature extraction on the metadata of each image by weighting low-rank adaptation (LoRA) experts within a frozen visual backbone. Across all eight species, MetaRouter outperforms visual-only and metadata-aware baselines. Controlled comparisons further show that the gains depend on using the metadata paired with each image to guide routing.

Primary area:
datasets and benchmarks
Keywords:
Animal Re-Identification, Wildlife Dataset, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_2LJWOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MetaFauna, an animal re-identification dataset containing 44,317 images of 2,500 individuals across eight species. It combines three existing data sources, adds identity annotations for Snapshot Safari, and provides structured metadata describing visibility, capture conditions, and environmental context. The paper also proposes MetaRouter, which uses metadata to route LoRA experts within a frozen DINOv3 backbone. Experiments across eight species, five training seeds, and controlled ablations show improvements over visual-only and metadata-aware baselines. Additional experiments examine missing metadata and joint multi-species training.

Critical strengths

MetaFauna provides image-level descriptions of which identity cues are visible and how clearly they are captured. This supports a useful research question that identity labels alone cannot directly address: how observation conditions should guide animal identification. The consistent schema and source mappings strengthen its value as a reusable benchmark. The attribution experiments are stronger than a simple comparison against external baselines. Static expert mixtures, visual routing, feature fusion, and shuffled metadata help separate the contributions of capacity, routing, and image–metadata correspondence. MetaRouter consistently improves over these controls across five seeds, providing meaningful evidence for metadata-conditioned adaptation.

Critical weaknesses

The new identity annotations lack a quantitative reliability assessment. Appendix A describes multiple annotators and cluster-level review, but reports neither agreement nor an independent audit of identity merges and splits. Reviewing likely matches can validate accepted associations while still missing links between separate identity clusters. Such missed links could place the same animal in nominally identity-disjoint partitions. This uncertainty directly affects the dataset’s central contribution and the validity of its evaluation. The protocols do not establish how well the benchmark measures identification across independent encounters. Snapshot Safari separates query and gallery by source sequence, but does not report their separation by camera site or capture time. Hyena and Leopard use image-level splitting because encounter information is unavailable. Near-duplicate or same-encounter matches therefore remain a possible shortcut. The reported gains establish performance on these partitions, but provide limited evidence of robustness across encounters.

Questions for authors
  1. What existing annotation-quality statistics are available, such as agreement rates, unresolved-pair rates, or independent identity audits? How were potentially missed associations outside the highest-ranked candidates checked?

  2. Were near-duplicate or same-encounter images audited across the Hyena and Leopard query/gallery partitions? Please summarize any existing checks.

  3. How many query identities in the Snapshot Safari subsets have gallery matches from different camera sites or substantially separated capture times? This would clarify the deployment setting represented by the benchmark.

6 comments

Thank you for the constructive review and for recognizing the value of MetaFauna's observation-level metadata schema and the controlled attribution study in Sec. 4.3. The concerns focus on the quantitative reliability of the new Snapshot Safari identity annotations, including potential false merges and missed associations across partitions, and the extent to which the evaluation splits prevent same-encounter shortcuts and evaluate identification across independent encounters. We address each point below through: (1) species-level inter-annotator agreement statistics, adjudication outcomes, and an independent audit of 270 identity clusters (Q1.1); (2) a calibrated retrieval-based audit of 5,781 cross-cluster candidate pairs confirming zero train–test identity conflicts among the reviewed pairs (Q1.2); (3) a similarity-aware audit of all 9,935 Hyena and Leopard images confirming that none of the 655 near-duplicate or same-encounter groups spans the query/gallery split (Q2); and (4) a spatial and temporal breakdown of query–gallery separation across the Snapshot Safari subsets (Q3).

1.1 What existing annotation-quality statistics are available, such as agreement rates, unresolved-pair rates, or independent identity audits?

Thank you for highlighting the importance of quantifying annotation reliability. Across the six Snapshot Safari species, initial inter-annotator agreement reaches 91.0% (κ=0.825\kappa=0.825), and an independent audit of 270 identity clusters confirmed zero incorrect identity merges. Below we provide additional agreement statistics and an independent identity audit for the six Snapshot Safari species whose individual identities were annotated in MetaFauna. Hyena and Leopard retain their source-provided identity labels and are therefore excluded from these analyses.

Annotation agreement. As described in Appendix A.1.2, two annotators independently classify candidate sequence-group pairs as same, different, or unresolved. A third annotator reviews all proposed positive associations and cases involving disagreement or uncertainty, and only same-individual associations confirmed by all three annotators are accepted.

Across 30,217 reviewed candidate pairs, the initial agreement rate is 91.0%, with Cohen's κ=0.825\kappa=0.825 and 91.9% positive agreement (PA) on same-individual decisions. The unresolved-pair rate is 4.4%, defined as the proportion of pairs for which at least one initial annotator selected unresolved. The species-level statistics are summarized below.

Inter-annotator agreement across Snapshot Safari subsets. Entries report pair-level agreement statistics prior to third-annotator adjudication.

SpeciesReviewed pairsAgreement (%)Cohen's κ\kappaPA on same (%)Unresolved (%)
Cheetah3,48790.40.80889.84.9
Elephant4,92688.30.77890.15.3
Giraffe6,43891.90.84493.34.2
Kudu4,07187.80.76387.95.6
Lion3,91390.60.81690.84.5
Zebra7,38294.10.88595.43.1
Overall30,21791.00.82591.94.4

In total, 14,770 candidate pairs underwent third-annotator review, comprising 11,880 pairs initially marked same by both annotators, 2,732 initial disagreements, and 158 pairs marked unresolved by both annotators. Following adjudication, 11,482 associations were confirmed as same, 2,784 were classified as different, and 504 remained unresolved. Unresolved pairs were never used to establish identity associations.

Independent identity audit. In addition to the cluster-level consistency checks performed during annotation, we independently examined 270 identity clusters (45 per species), covering 7,140 images and 2,720 source sequences. These audited clusters include identities with previously disputed associations, large clusters, and identities spanning different cameras or capture dates, alongside randomly selected remaining identities. Independent reviewers inspected the complete sequences and persistent individual-specific visual features without access to the original similarity scores or identity decisions. No incorrect identity associations were confirmed among the audited clusters. We have added the agreement statistics, adjudication outcomes, and independent audit details to Appendix A.1.3 (Table A.1) of the revised PDF.

1.2 How were potentially missed associations outside the highest-ranked candidates checked?

An independent retrieval audit of 5,781 cross-cluster candidate pairs confirmed zero train–test identity conflicts and resolved two missed test-partition associations prior to finalizing the benchmark. Specifically, during dataset construction, we used an independent visual retrieval model to identify potential same-individual associations overlooked by the original DINOv3 candidate ranking, followed by manual verification of the corresponding source sequences.

Retrieval calibration. We first evaluated whether the independent model could recover known associations that received low DINOv3 rankings. In a leave-one-encounter-out evaluation across 286 eligible cases, the model recovered the correct identity within its top-5 candidates in 245 cases (85.7%). Among the 20% of cases ranked lowest by DINOv3 (58 cases), it recovered 41 (70.7%), showing its ability to retrieve challenging associations that received low rankings from the original model.

Independent candidate audit. If a true association was missed during initial candidate screening, the same individual would remain split across two separate identity clusters. To identify such cases, we used the independent model (k=5k=5) to retrieve additional candidate matches within and across the provisional training and test partitions, yielding 5,781 unique candidate pairs. All retrieved candidate pairs were manually reviewed using the complete source sequences and the original verification criteria. No train–test identity conflicts were confirmed among the reviewed pairs, while two missed associations were identified within the test partition and resolved before finalizing the identity mapping and evaluation splits. The dataset statistics, fixed splits, and experimental results reported in the submission correspond to this finalized benchmark, and we have documented this retrieval audit in Appendix A.1.3 of the revised PDF.

2. Were near-duplicate or same-encounter images audited across the Hyena and Leopard query/gallery partitions?

Thank you for raising this concern. Our similarity-aware audit across all 9,935 Hyena and Leopard images confirmed 655 near-duplicate or same-encounter groups, with zero confirmed groups spanning the query/gallery partitions. MetaFauna adopts identity-disjoint train/test partitions to evaluate generalization to unseen individuals. For the six Snapshot Safari species, query and gallery images are separated at the source-sequence level. Because HyenaID2022 and LeopardID2022 do not provide encounter identifiers for direct encounter-level splitting, we incorporated similarity-aware verification during split construction to prevent repeated observations from creating easy positive matches across their query/gallery partitions.

Similarity-aware verification. We examined all 9,935 Hyena and Leopard images using exact-duplicate checks, perceptual similarity across both animal crops and original frames (capturing near-identical images that differ in cropping, encoding, or minor visual details), and the similarity-aware visual clustering procedure from WildlifeReID-10k [1]. Candidate pairs were independently examined by two annotators, with a third annotator adjudicating positive, conflicting, or uncertain cases. Near-duplicates were identified based on substantially repeated visual content, while same-encounter decisions additionally considered available camera identifiers, timestamps, burst information, and scene consistency.

Audit results and split validation. Across the 9,935 images from the two source datasets, the WildlifeReID-10k visual clusters mapped to MetaFauna identify 727 candidate visual groups involving 1,982 images from 174 identities (Hyena: 132 groups, 322 images, 55 identities; Leopard: 595 groups, 1,660 images, 119 identities). Combining these groups with additional perceptual-similarity candidates, independent human review confirmed 655 near-duplicate or same-encounter groups, comprising 1,743 images from 162 identities. For test identities, each confirmed repeated-observation group was assigned entirely to either query or gallery while preserving valid positive matches. A final check found no confirmed groups spanning both partitions. We have included these similarity-aware audit and split validation details in Appendix C.1 (and the foreground-only evaluation in Appendix F.2, Table F.3) of the revised PDF.

[1]. Adam, Lukas, et al. "WildlifeReID-10k: Wildlife re-identification dataset with 10k individual animals." Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.

Reviewer_w2fdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

this paper introduces metafauna. an animal ReID dataset provides structured metadata describing visual evidence visibility, capture conditions, and environmental context. A method called MetaRouter is also desinged to use LoRA experts within a frozen visual backbone, allowing feature adaptation to vary with the metadata of each image. Experiments demonstrate improvements over existing animal ReID baselines.

Critical strengths

1、This work provides a reid benchmark, which may be a potentially valuable dataset conrtibution to boost the research in reid fields. 2、Comprehensive evaluation and analysis are made on the benchmarks. 3、This work analyze the expert miextures, routing and metadata which provide evidences that correct uses of metadata are beneficial to the final performance.

Critical weaknesses

1、The method of MetaRouter is limited novelty. It seems like simply applying the already established mechanisms to animal reid rather than carefully designing a new fundamentally method. 2、The quality of the annotation can be more provided and analayzed for a dataset work, how to solve the potenial error introduced throughout the whole data processing procedure. 3、Regarding to results, different backbones are used in the experiments. The necessity of isolation of the architectureal improvements should be provided. 4、The generalization of proposed methods should be analyzed not specific to the proposed datasets, and more results should be added.

Questions for authors

1、Does MetaRouter generalize to external ReID datasets? 2、How much performace are corrected after automatically predicted? 3、can the author provide more fair comparions of metarouter agaist other baselines?

8 comments

Thank you for the constructive review and for recognizing MetaFauna's value as an animal ReID benchmark, our comprehensive experimental analysis, and the evidence showing how structured metadata and routing improve retrieval. The questions and concerns focus on generalization to external ReID datasets, performance with automatically predicted metadata, fair comparisons that isolate MetaRouter's architectural contributions from backbone differences, the positioning and novelty of MetaRouter, and quality control across the dataset construction pipeline. We address each point below through: (1) external evaluations on MetaWild and CzechLynx across seven additional species (Q1 / Weakness 4); (2) inference experiments using automatically predicted and corrupted G1 metadata (Q2); (3) controlled same-backbone attribution experiments across DINOv3 and SigLIP 2 architectures (Q3 / Weakness 3); (4) clarifying MetaRouter's role as a controlled reference framework for metadata-guided backbone adaptation (Weakness 1); and (5) quantitative audits and error-control checks across all stages of dataset construction (Weakness 2).

Q1 (Weakness 4). Does MetaRouter generalize to external ReID datasets?

We thank the reviewer for raising this question. We also appreciate the contributions of researchers and field teams who have made annotated wildlife datasets available to the community. MetaRouter generalizes effectively beyond MetaFauna, achieving the highest retrieval accuracy across all seven external species evaluated. Specifically, to investigate whether MetaRouter's benefits extend beyond MetaFauna, we evaluate it on two additional Animal ReID datasets: MetaWild [1], which provides environmental metadata for six species (Deer, Hare, Penguin, Pūkeko, Stoat, and Wallaby), and CzechLynx [2], whose authors provide detailed camera-trap records, including trap identifiers, encounter annotations, observation dates, and coarse geographic information.

Experimental protocol. We train each method independently on the training identities of each external dataset and evaluate on unseen test identities. For MetaWild, we follow the official train/query/gallery partitions. For CzechLynx, we adopt the official Geo-aware train/test split, comprising 21,763 training images from 224 individuals and 17,997 test images from 95 unseen individuals. We further construct fixed, date- and encounter-separated query/gallery partitions containing 2,761 queries and 15,215 gallery images, covering 93 eligible test identities.

Metadata configuration. MetaRouter supports diverse metadata fields without requiring the full MetaFauna G1/G2/G3 schema. For MetaWild, we use its original face direction, temperature, and day/night attributes, combined with image-derived capture conditions. For CzechLynx, we use automatically predicted G1 annotations, image-derived G2 attributes, and environmental context derived from available observation dates and coarse spatial records. Fields that cannot be reliably obtained remain masked. We retain the original MetaRouter architecture and training objectives, adjusting only the dataset-specific metadata inputs.

Comparison methods. We compare MetaRouter against CLIP-ReID+MFA and ReID-AW+MFA, two established metadata-aware ReID methods, as well as Shared LoRA, Static Expert Mix, and Shuffled MetaRouter. All methods use identical dataset partitions, with metadata-aware methods receiving the same available metadata fields. The LoRA-based controls share MetaRouter's DINOv3 backbone and adaptation settings, allowing us to isolate the contributions of expert capacity and metadata-conditioned routing. Results are averaged over five training seeds.

Generalization to external Animal ReID datasets. Each entry reports mAP / Rank-1 / Rank-5 (%). Higher is better.

DatasetSpeciesCLIP-ReID+MFAReID-AW+MFAShared LoRAStatic Expert MixShuffled MetaRouterMetaRouter
MetaWildDeer70.3/97.7/99.571.9/97.0/99.671.5/97.2/99.472.4/97.5/99.770.4/96.8/99.375.8/98.4/99.7
Hare62.8/95.1/98.867.1/97.1/99.265.3/96.4/99.164.9/96.2/99.065.1/96.1/98.968.2/97.5/99.2
Penguin51.7/70.0/88.356.6/72.4/89.853.1/70.8/88.454.2/71.3/89.153.6/70.9/88.758.7/75.6/92.4
Pūkeko60.9/84.5/95.162.9/86.4/96.163.4/85.2/95.562.1/85.8/95.462.5/84.6/95.265.3/87.3/96.8
Stoat72.4/92.6/98.374.8/95.4/98.972.6/94.0/98.573.5/94.5/98.772.3/93.6/98.377.9/96.1/99.4
Wallaby63.4/93.1/98.164.2/93.4/98.362.6/92.6/97.763.9/92.8/98.263.1/92.7/97.966.1/94.1/98.6
CzechLynxEurasian Lynx10.4/31.8/45.611.7/34.4/49.311.2/33.1/47.711.9/33.7/48.610.7/32.4/47.214.6/39.3/54.7

Results. MetaRouter achieves the highest mAP across all seven external species. Compared with Shared LoRA and Static Expert Mix, it improves mAP by an average of 3.84 and 3.39 points, respectively. Shuffled MetaRouter performs comparably to the metadata-independent controls, while MetaRouter achieves an average improvement of 4.13 mAP points over the shuffled variant. These comparisons show that the improvements cannot be attributed to additional expert capacity alone, but also depend on the correspondence between each image and its metadata.

MetaRouter also outperforms ReID-AW+MFA by 2.49 mAP points on average, with consistent improvements across the six MetaWild species. On CzechLynx, MetaRouter achieves 14.6% mAP, 39.3% Rank-1, and 54.7% Rank-5, compared with 11.2%, 33.1%, and 47.7% for Shared LoRA. Although absolute retrieval accuracy is lower on CzechLynx, the evaluation involves geographically distinct training and test populations, unseen identities, and date-separated query/gallery observations. The improvements over the same-backbone controls provide further evidence for the usefulness of metadata-conditioned adaptation beyond MetaFauna. We have included these external dataset evaluations in Appendix F.3 (Table F.4) of the revised PDF.

[1]. Li, Yuzhuo, et al. "Metawild: A multimodal dataset for animal re-identification with environmental metadata." Proceedings of the 33rd ACM international conference on multimedia. 2025.

[2]. Picek, Lukas, et al. "Czechlynx: A dataset for individual identification and pose estimation of the eurasian lynx." Scientific Data 13.1 (2026): 511.

Q2. How much performance is retained with automatically predicted metadata?

With automatically predicted G1 metadata at inference, MetaRouter achieves 66.93% mean mAP, retaining approximately 69% (+2.63 points) of the verified-metadata gain over Shared LoRA. To quantify this effect, we replace manually verified G1 annotations at inference with predictions while retaining the original MetaRouter checkpoints, G2/G3 metadata, and evaluation splits. Across eight species, mean mAP decreases from 68.09% with verified G1 to 66.93% with predicted G1 (a reduction of 1.16 points), while remaining 2.63 points above that visual-only control (64.30%).

Automatic G1 prediction. We train an independent visual predictor for viewpoint, visible regions, and external occlusion using training identities only. Boundary contact is derived automatically from image geometry and detection boxes. At evaluation, the predicted G1 fields replace the verified values for both query and gallery images. The predictor achieves 0.887 macro-F1 for viewpoint, 0.916 macro-F1 for visible regions, 93.2% balanced accuracy for external occlusion, and 0.973 macro-F1 for boundary contact.

Performance with automatically predicted G1 metadata (mAP, %).

SpeciesShared LoRAVerified G1Predicted G1
Cheetah49.154.252.5
Elephant70.572.071.9
Giraffe77.179.278.7
Hyena55.361.559.6
Kudu76.980.079.6
Leopard46.451.649.2
Lion69.972.672.1
Zebra69.273.671.8
Mean64.3068.0966.93

The performance reduction varies across species, with larger decreases on Leopard, Hyena, Zebra, and Cheetah, while MetaRouter with predicted G1 remains above Shared LoRA on all eight species.

Robustness to inaccurate metadata. In a separate controlled evaluation, corrupting 20% of available G1 fields decreases mean mAP by only 0.46 points relative to verified G1 (67.63% vs. 67.56% when the same fields are removed), indicating low sensitivity to moderate errors. At 60% corruption, incorrect values yield 64.79% mAP versus 66.76% when the same fields are removed, and fully corrupted G1 reduces mAP to 61.84% (compared with 66.07% when removed), showing that severely inaccurate metadata is substantially more harmful than missing metadata. To keep this response concise, we provide the full 0%–100% corruption progression table in our response to Reviewer 13T8 (Q5) and have included both analyses in Appendix F.1 (Tables F.1–F.2) of the revised PDF.

Reviewer_13T8OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper contributes meta-annotations to existing publicly available camera-traping data. The authors generated annotations about viewpoint, visible body region, image quality, environmental conditions, etc. for around 44k images across eight species. The authors also introduce MetaRouter, a model that uses this metadata based on LoRA and DINOv3. The model outperforms a set of baselines across all eight species, and the authors provide several ablations to analyze the improvements.

Critical strengths

The main strength of the paper is the new meta-annotations that could be a useful resource for further animal ReID research.

Critical weaknesses

My main concern is related to the dataset construction and reliability of the provided annotations. I do not really understand why the authors decided to annotate individual identities instead of extending existing animal ReID datasets, e.g., those available through WildlifeDatasets which would be, in my opinion, a much more useful contribution to the community. I do not understand why the authors chose to create new identity labels instead of using datasets where these labels already exist, as they did for Hyenas and Leopards. Besides, identifying individual animals is extremely difficult and prone to errors. Even experts studying one species for their life can have +10 % error rate. Without quantitative evidence of the annotation quality, I find it hard to trust the new labels.

I am also not sure how useful MetaFauna would be as a benchmark. The authors state that they will release the annotations and mappings to the original images, but the images themselves come from several different sources and are not redistributed. I assume one of the reasons WD was developed was to make access and evaluation simple. Having to reconstruct MetaFauna from multiple sources, match the annotations, and reproduce the splits seems pretty complicated.

Another concern is potential information leakage through the metadata. Both Hyena and Leopard datasets have strong background priors, and therefore metadata are correlated with identities. Using such metadata might therefore improve performance simply by leaking the background information. This is in my opinion potentially harmful given that the query/gallery split for these datasets is done at the image level, i.e. randomly. Given that, I am not convinced that the reported improvements necessarily demonstrate better animal ReID.

I also have concerns about the practical usefulness of MetaRouter. Some metadata can be extracted automatically, but many attributes need manual annotation. Where would this metadata come from in a real-world deployment? The authors investigate missing metadata, but not automatically predicted and potentially inaccurate metadata. Without this experiment, I cannot tell how much of the improvement could we get in practice.

Questions for authors
  • Q1 Why did you choose to create new identity labels instead of extending existing ReID datasets?
  • Q2 How did you measure the reliability of the identity and metadata annotations?
  • Q3 Could you disprove that the on Hyena and Leopard comes from background or train-->test leakage?
  • Q4 How will researchers access MetaFauna and reproduce the evaluation splits?
  • Q5 How well does MetaRouter perform with predicted / inaccurate metadata?
  • Q6 Does MetaRouter generalize to datasets outside MetaFauna?
  • Q7 How are the reviewers supposed to assess you really build the "dataset" without having access to the dataset? Imho, just a promise "we will release it once accepted" should not be tolerated in case of dataset contributions.
8 comments

Thank you for the constructive review and for recognizing the potential of MetaFauna's metadata annotations for animal ReID research. The comments center on dataset construction and annotation reliability, benchmark accessibility and split integrity, and practical deployment. We address these concerns through: (1) clarifying the design rationale for combining existing ReID datasets with Snapshot Safari identities and reporting quantitative verification of identity and G1–G3 metadata annotations (Q1–Q2); (2) verifying same-encounter split integrity and evaluating foreground-only controls on Hyena and Leopard (Q3), alongside providing the complete annotation package and single-command reconstruction tool (Q4, Q7); and (3) evaluating inference with automatically predicted or corrupted metadata and testing generalization across seven external species on MetaWild and CzechLynx (Q5–Q6).

Q1. Why did you choose to create new identity labels instead of extending existing ReID datasets?

Thank you for asking about our decision to annotate new identities. MetaFauna both extends existing identity-labeled datasets and establishes new identities on Snapshot Safari so that G1, G2, and G3 metadata can be studied under a unified multi-species schema. We greatly value the contributions of existing public animal ReID datasets, including those accessible through WildlifeDatasets [1] and WildlifeReID-10k [2]. These resources provide a valuable foundation for animal ReID research. MetaFauna also builds on existing identity annotations: for Hyena ID 2022 and Leopard ID 2022, we retain the original identity labels and extend these datasets with structured metadata.

For the remaining six species, our choice of Snapshot Safari was motivated by the source information required to construct structured observation metadata. Existing identity-labeled datasets vary in the availability of original frames, camera locations, sequence information, and capture timestamps. Although some provide these records, their availability varies across datasets, making it difficult to define the same metadata fields across multiple species. Snapshot Safari provides original camera-trap images alongside source sequences, camera locations, and capture timestamps. These records allow us to annotate which identity-relevant visual cues are visible (G1), characterize how they are captured (G2), and derive environmental context (G3). However, Snapshot Safari’s original annotations are primarily at the species level, making additional individual identity annotation necessary for ReID. We therefore established individual identities for 34,382 images of 1,814 individuals across six species, providing a consistent metadata schema while preserving the original capture information. While MetaFauna provides this unified benchmark for studying G1–G3 metadata, our evaluations on MetaWild and CzechLynx (Q6) further show that MetaRouter adapts effectively to existing public datasets with partial, dataset-specific metadata. We hope this complementary contribution will provide a useful resource for the community alongside existing public ReID datasets.

[1]. Čermák, Vojtěch, et al. "Wildlifedatasets: An open-source toolkit for animal re-identification." 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2024.

[2]. Adam, Lukas, et al. "WildlifeReID-10k: Wildlife re-identification dataset with 10k individual animals." Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.

Q2. How did you measure the reliability of the identity and metadata annotations?

We assessed annotation reliability through independent human verification of identity and visual metadata annotations, together with computational and source-level validation of the remaining metadata fields.

Identity annotation reliability. For the six Snapshot Safari species, the initial agreement across 30,217 candidate pairs was 91.0% (Cohen's κ=0.825\kappa=0.825), with 4.4% involving an unresolved decision. An independent audit of 270 identity clusters, covering 7,140 images and 2,720 source sequences, identified no incorrect identity associations among the audited clusters. To identify potentially missed associations beyond the original DINOv3 candidate ranking, we additionally used an independent retrieval model to propose candidate matches between separate identity clusters, followed by manual verification of the complete source sequences. No train–test identity conflicts were confirmed among the reviewed pairs, while two missed associations between provisional test clusters were confirmed and resolved before the final identity mapping and evaluation splits were established.

Visual metadata reliability (G1). As part of dataset construction, we verified G1 annotations across all 44,317 images through independent dual annotation and third-annotator adjudication. Two annotators assessed viewpoint, visible body regions, external occlusion, and boundary contact using the original frames and verified target boxes. A third annotator reviewed disagreements, uncertain cases, and a random subset of agreed annotations to establish the final verified labels. The pre-adjudication reliability statistics are summarized below.

Pre-adjudication reliability of G1 visual metadata across all 44,317 images.

G1 fieldHuman agreementReliability metricInitial annotations vs. verified labels
Viewpoint95.8%κ=0.928\kappa=0.92896.2% selective accuracy
Visible regions97.4%Macro-F1 = 0.971Macro-F1 = 0.966
External occlusion96.6%κ=0.913\kappa=0.91396.3% balanced accuracy
Boundary contact98.1%Macro-F1 = 0.95697.1% balanced accuracy

The final annotations incorporate the verified corrections, while ambiguous observations remain marked as unknown. We have provided the detailed agreement statistics, field-level coverage, and annotation guidelines in Appendix A.2.1–A.2.2 (Table A.3) of the revised PDF.

Computed and source-derived metadata (G2–G3). Since G2 and G3 primarily involve deterministic computations or external-data matching, we evaluate their reliability through independent recomputation and source consistency checks. For G2, brightness and sharpness were reproduced across all 44,317 images within numerical precision, while target scale was independently checked against the original-frame geometry. Imaging-mode assignments were also examined against source-frame appearance and available capture information. For G3, independently calculated solar-period labels matched all 34,230 available annotations, and temperature and precipitation values were reproduced for all 34,212 source-matched observations. We additionally examined timestamp consistency, camera-location matching, and the spatial resolution of external environmental records. Fields without sufficiently reliable source information remain missing. Detailed audit protocols, field-level statistics, and source-validation procedures have been included in Appendix A.1.3 (Table A.1), A.2.2 (Table A.3), and A.2.5 of the revised PDF.

Q3. Could the improvements on Hyena and Leopard arise from background information or train–test leakage?

Thank you for raising this concern. Our split verification and foreground-only experiments show that MetaRouter's gains on Hyena and Leopard persist when repeated observations are kept within the same partition and visible image backgrounds are removed. We address split integrity, metadata scope, and background dependence below.

Split integrity and repeated observations. MetaFauna adopts identity-disjoint training and test partitions, ensuring that no test individual appears during training. For Hyena and Leopard, where reliable encounter identifiers are unavailable, we account for repeated observations through similarity-aware verification, following the principle adopted in WildlifeReID-10k [2]. Specifically, we screened all 9,935 Hyena and Leopard images using perceptual similarity across animal crops and original frames alongside the visual cluster assignments released with WildlifeReID-10k (727 candidate visual groups), followed by manual verification of candidate groups and potentially conflicting identity associations. Across all 9,935 images, independent verification confirmed 655 near-duplicate or same-encounter groups involving 1,743 images from 162 identities. Confirmed repeated observations are kept within the same query/gallery partition, while cross-identity candidates are examined to prevent identity overlap between training and test sets. None of the 655 confirmed groups spans the reported query/gallery partitions.

Metadata scope and conditioning mechanism. Furthermore, the metadata used for Hyena and Leopard does not include G3 environmental records, camera locations, or capture timestamps; all G3 fields are unavailable for both species. Only G1 visual attributes (viewpoint, visible body regions, external occlusion, and boundary contact) and G2 capture attributes (brightness, sharpness, target scale, and imaging mode) are provided. MetaRouter uses these attributes to condition visual feature extraction within the backbone, without directly including metadata in the retrieval similarity computation. Because capture conditions or viewpoints could still correlate with specific camera traps, we directly test whether the resulting improvements rely on visual background cues in the experiments below.

Background dependence. To examine whether MetaRouter's improvement relies primarily on background information, we conduct foreground-only and background-only evaluations. In the foreground-only setting, image backgrounds are removed using verified animal masks, and all compared methods are trained and evaluated using the same data partitions and optimization settings. The original observation metadata is retained to isolate the influence of visual background information. We compare MetaRouter with Shared LoRA and Shuffled MetaRouter, which preserves the same architecture while disrupting image–metadata correspondence. The results are summarized below (mAP, %).

Full-image and foreground-only evaluation on Hyena and Leopard (mAP, %).

SpeciesSettingShared LoRAShuffled MetaRouterMetaRouter
HyenaFull-image55.354.861.5
Foreground-only51.650.956.8
LeopardFull-image46.446.551.6
Foreground-only41.842.446.7

MetaRouter retains improvements of 5.9 and 4.3 mAP over Shuffled MetaRouter under foreground-only evaluation on Hyena and Leopard, respectively, compared with 6.7 and 5.1 mAP using full images (and improves over Shared LoRA by 5.2 and 4.9 mAP with foreground-only inputs). We further evaluate a background-only visual baseline without metadata, obtaining 11.4% and 15.2% mAP, respectively. These results indicate that background information provides some identity-correlated signal but is insufficient to explain the performance of the proposed framework. MetaRouter's gain over Shuffled MetaRouter remains under the verified partitions and with foreground-only inputs, so it cannot be attributed only to repeated observations or background cues. We have added the split verification details to Appendix C.1 and the foreground-only evaluation results to Appendix F.2 (Table F.3) of the revised PDF.

Reviewer_v9UvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a new Animal Re-ID dataset, MetaFauna, which includes 44,317 images across 2500 individuals spanning 8 different species. In addition to the dataset, a new model, MetaRouter, is proposed which leverages both visual data and meta-data (visual evidence visibility, capture conditions, and environmental context) for improved recognition (re-id) accuracy. Experiments are performed to demonstrate both that the metadata-wielding MetaRouter model outperforms both visual-only and metadata-aware baselines, and, via ablations, where the benefit comes from.

Critical strengths
  • Clarity/Motivation: The paper is well-written and clearly motivates the need for/value of a model that can leverage both the visual image information and relevant meta-data, particularly if it can delineate identity-related differences from visual variation in observations.
  • Experiments: The experiments are well selected to demonstrate the effectiveness of a model that incorporates relevant metadata. The supplementary material/appendix is very thorough, both in the additional experiments it provides, and in the details of the dataset collection, curation and composition.

Notes:

  • I also found Figure 2 a very helpful way to describe the dataset density/distribution over species/images.
Critical weaknesses
  • Limited Novelty: There is limited novelty in the model. The model is effectively a baseline leveraging a weighted mixture of LoRA experts who incorporate meta-data. While there is limited novelty in the approach it does perform better than some other (e.g. CLIP-based) baselines. It also feels like it stops a bit short in showing fine-grained impacts of the different types of metadata (the fields as referred to in the appendix).
  • Experiments: The experiments are well done in the range of things they cover and show. However, at least for me, the limited novelty of the MetaRouter baseline bring into question the convincingness of the experiments. They do a good job demonstrating that metadata helps the model to perform better, but don't answer the question of what might be possible when combining meta-data with a very strong visual model.
Questions for authors

Clarifications:

  • It seems that the metadata dropout is used during training and during inference (L323) the dropout is disabled, but both query + gallery images are encoded with available metadata. Is it possible to leverage the metadata during training and to use visual data only at inference time? Is this the same as the p=1 meta-data missing setting experiment?
  • Also, you have shown the relative impact of G1, G2 and G3... I wondered if you could similarly break out the impact of the individual measures that compose each of these. I think this is what is in Figure D.1, is that correct? If so, I feel that this would be highly valuable to include in the main paper, rather than supplementary.

Suggestions:

  • It felt redundant to have the same quantitative info on the dataset several times (L24, L88, L163) and similarly the characterizations of G1, G2, G3 several times (Figure 1, L90-95, L238-242).
4 comments

Thank you for the constructive review. We appreciate your positive assessment of the paper's motivation, experimental design, and detailed supplementary analyses. The questions and concerns focus on inference without metadata, the fine-grained contributions of individual metadata fields, the positioning of MetaRouter, and whether structured metadata remains beneficial when paired with stronger visual backbones. We address these points through: (1) clarifying visual-only inference and its exact equivalence to the p=1p=1 setting in original Fig. 5 (now Fig. 6 in the revised PDF; Q1); (2) summarizing the field-level retraining ablations from original Appendix Fig. D.1 and moving them into the revised main paper as Fig. 5 (with details in Appendix D.3) while streamlining repeated dataset and G1–G3 descriptions (Q2, Suggestions, and W1); (3) clarifying MetaRouter's role as a controlled reference framework under a shared visual backbone (W1); and (4) adding a backbone-scaling and cross-paradigm study across DINOv3 (ViT-B/16 to ViT-7B/16) and SigLIP 2 So400M in Table 4 and Appendix F.4 (W2).

Q1. Can metadata be used only during training, with visual-only inference? Is this equivalent to the p=1p=1 metadata-missing setting?

Yes. MetaRouter supports training with metadata while requiring only images at inference. This is the p=1p=1 setting in original Fig. 5 (now Fig. 6 in the revised PDF; Appendix D.4), where all available metadata fields are removed from both query and gallery images at inference, without changing the trained model.

During training, field-level and full-record metadata dropout expose MetaRouter to partially or entirely missing metadata. When no metadata is available at inference (Fi=∅\mathcal{F}_i=\emptyset), the metadata-conditioned expert routing is disabled, and visual features are extracted using the shared LoRA path (Sec. 4; Appendix C.1).

As shown in revised Fig. 6 (original Fig. 5), MetaRouter at p=1p=1 remains within 0.4 mAP points of the visual-only Shared LoRA baseline across all eight species. This shows that MetaRouter can operate without inference-time metadata, although the full benefit of metadata-conditioned adaptation is not retained. With metadata available, MetaRouter achieves 68.09% mean mAP, compared with 64.30% for Shared LoRA. These results distinguish the ability to perform metadata-free inference from the additional improvements obtained when observation metadata is available.

Q2. Can the contributions of individual metadata fields be analyzed beyond G1, G2, and G3?

Thank you for highlighting the value of a more fine-grained analysis and for suggesting its inclusion in the main paper. Yes, the analysis in original Appendix Fig. D.1 evaluates all individual metadata fields in G1, G2, and G3, and following your suggestion we have moved it into the revised main paper as Fig. 5 (with full details in Appendix D.3). We use two complementary settings: field-only, which measures the improvement from using a single field relative to the metadata-free baseline, and leave-one-field-out, which measures the performance decrease when a field is removed while all remaining metadata is retained. Both settings use the same data splits and training protocol.

Field-level metadata contributions. Entries report mean mAP difference in percentage points, averaged over applicable species.

GroupFieldField-onlyLeave-one-out
G1Viewpoint+2.5+0.9
G1Visible regions+2.5+1.0
G2Imaging mode+2.1+0.6
G2Sharpness+1.5+0.5
G3Solar period+1.0+0.4
G3Ecological season+0.7+0.2

Within G1, viewpoint and visible regions provide the largest contributions, while imaging mode and sharpness are the most informative G2 fields. G3 fields have smaller individual effects. Across all groups, leave-one-field-out drops are smaller than single-field gains, indicating that multiple fields capture partially overlapping observation cues instead of relying on any single attribute. Complete species-level breakdowns are provided in Fig. 5 and Appendix D.3 of the revised PDF.

To accommodate this analysis in the revised main paper (Fig. 5) while retaining the complete discussion in Appendix D.3, we have also streamlined the repeated dataset statistics and descriptions of G1–G3 in the Introduction and Dataset sections of the updated PDF.

Weakness 1. Limited Method Novelty and Fine-grained Metadata Analysis

Thank you for raising these two points. MetaFauna is primarily a dataset contribution, and MetaRouter serves as a controlled reference framework to study metadata-guided feature extraction inside a visual backbone.

Contribution and method positioning. Specifically, MetaFauna is designed to support systematic investigation of structured metadata in Animal ReID, while MetaRouter addresses a focused research question: can metadata describing observation conditions guide identity-discriminative feature extraction inside a visual backbone, beyond late feature-level fusion? Building on established LoRA and expert-routing mechanisms makes performance changes easier to attribute to how metadata is used, without introducing extra architectural modules whose effects would be difficult to isolate.

Beyond comparisons with CLIP-based baselines, Sec. 4.3 includes controlled experiments using the same pretrained DINOv3 [1] backbone and adaptation settings. Static Expert Mix (64.66% mAP) isolates the effect of additional expert capacity, while Feature Fusion (66.28%) and Visual Router (66.10%) examine alternative ways of incorporating metadata or determining routing weights. Shuffled MetaRouter (64.43%) further tests the importance of correctly paired image–metadata records. MetaRouter achieves 68.09% mean mAP, supporting the benefit of using image-specific metadata to guide visual adaptation inside the backbone.

Fine-grained metadata contributions. As summarized in our response to Q2, the field-level analysis (original Appendix Fig. D.1, now moved to Fig. 5 in the revised main paper and Appendix D.3) quantifies the individual impact of each attribute within G1, G2, and G3. Viewpoint (+2.5 field-only / +0.9 leave-one-out) and visible regions (+2.5 / +1.0) drive the largest gains in G1, while imaging mode (+2.1 / +0.6) and sharpness (+1.5 / +0.5) lead in G2. Moving this analysis into the revised main paper makes the fine-grained effects of individual metadata attributes directly visible alongside our group-level ablations.

Weakness 2. The value of metadata with stronger visual models

Thank you for raising this important question. Structured metadata remains consistently beneficial as visual backbones scale, with MetaRouter improving mean mAP by +2.55 to +4.12 points over Shared LoRA across larger DINOv3 models and SigLIP 2. To examine how visual backbone strength affects the utility of structured metadata, we first highlight our existing controlled comparisons on DINOv3 and then extend the evaluation across larger DINOv3 backbones and a vision-language pretrained architecture.

Existing strong visual baselines. Our main experiments already build on a pretrained DINOv3 ViT-B/16 backbone. In Sec. 4.3, the visual-only Shared LoRA control achieves 64.30% mean mAP, outperforming CLIP-ReID (59.59%) and matching AdaFreq (64.34%), the strongest visual-only baseline in Table 1. On the same DINOv3 backbone, MetaRouter reaches 68.09% (+3.79 points), whereas Shuffled MetaRouter obtains 64.43%. These results show that the gain from paired metadata holds even when the visual-only baseline is already competitive with state-of-the-art animal ReID methods.

Backbone scaling and cross-paradigm evaluation. To test whether metadata remains beneficial as visual representations become stronger, we evaluate three larger DINOv3 backbones: ViT-L/16 (300M), ViT-H+/16 (840M), and ViT-7B/16 (6.7B), and SigLIP 2 So400M/16-256 [2] (400M), which represents a distinct vision-language pretraining paradigm. Across all eight species and five training seeds, we compare Shared LoRA, Static Expert Mix, Feature Fusion, Shuffled MetaRouter, and MetaRouter using identical data partitions and matched adaptation settings within each backbone.

Backbone-scaling and cross-paradigm evaluation on MetaFauna. Entries report eight-species mean mAP / Rank-1 (%).

BackboneShared LoRAStatic Expert MixFeature FusionShuffled MetaRouterMetaRouter
DINOv3 ViT-B/16 (86M)64.30 / 77.8464.66 / 77.9266.28 / 78.9664.43 / 77.7168.09 / 79.78
DINOv3 ViT-L/16 (300M)67.92 / 80.5168.27 / 80.7669.66 / 81.7368.08 / 80.3971.34 / 82.91
DINOv3 ViT-H+/16 (840M)70.38 / 82.2470.61 / 82.4371.79 / 83.2170.25 / 82.1673.27 / 84.28
DINOv3 ViT-7B/16 (6.7B)71.36 / 82.9371.69 / 83.1272.59 / 83.7471.48 / 82.8573.91 / 84.67
SigLIP 2 So400M (400M)65.84 / 78.8666.19 / 79.1468.03 / 80.5265.71 / 78.7969.96 / 81.83

Results and analysis. Scaling DINOv3 from ViT-B/16 (86M) to ViT-7B/16 (6.7B) increases visual-only Shared LoRA performance from 64.30% to 71.36% mean mAP (and 77.84% to 82.93% Rank-1), with substantial gains from ViT-B to ViT-L (+3.62 mAP) and ViT-H+ (+2.46 mAP) before moderating at ViT-7B (+0.98 mAP). Notably, ViT-B/16 with MetaRouter (68.09% mAP) outperforms visual-only Shared LoRA on both ViT-L/16 (67.92%) and SigLIP 2 So400M (65.84%), showing that explicit observation metadata enables an 86M backbone to surpass visual-only models 3.5–4.6× its size. As visual capacity increases, the relative gain from metadata naturally moderates as stronger backbones resolve more visual ambiguities, yet combining MetaRouter with larger backbones still pushes performance to 71.34% on ViT-L/16 (+3.42 mAP), 73.27% on ViT-H+/16 (+2.89 mAP), and 73.91% on ViT-7B/16 (+2.55 mAP, 84.67% Rank-1).

The controlled comparisons show that Static Expert Mix and Shuffled MetaRouter remain within 0.4 mAP points of Shared LoRA across all four DINOv3 backbones, confirming that additional expert capacity and shuffled metadata alone do not account for the improvement. Feature Fusion provides consistent gains, while MetaRouter further outperforms Feature Fusion by 1.68, 1.48, and 1.32 mAP points on ViT-L, ViT-H+, and ViT-7B, respectively (with Rank-1 gains of 0.93–1.18 points). These comparisons support the benefit of using correctly paired metadata to guide visual adaptation inside the backbone, beyond late feature-level fusion.

On SigLIP 2 So400M, MetaRouter achieves 69.96% mAP and 81.83% Rank-1, improving over Shared LoRA by 4.12 mAP points and over Feature Fusion by 1.93 points. This extends the evaluation beyond DINOv3 and shows that metadata-conditioned adaptation is particularly effective for vision-language pretrained representations. We have included a concise summary table in the revised main paper (Table 4) and the full backbone-scaling and cross-paradigm results in Appendix F.4 (Table F.5) of the revised PDF.

[1]. Siméoni, Oriane, et al. "Dinov3." arXiv preprint arXiv:2508.10104 (2025).

[2]. Tschannen, Michael, et al. "Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features." arXiv preprint arXiv:2502.14786 (2025).

Thank you again for your constructive review and helpful suggestions. As the discussion period comes to a close, we would like to briefly follow up to see if our response and the revised PDF have addressed your questions, particularly regarding visual-only inference (p=1), moving the field-level metadata analysis into the main paper (now Fig. 5), and scaling MetaRouter up to DINOv3 ViT-7B and SigLIP 2 (Table 4 and Appendix F.4). Please let us know if any point remains unclear or if further clarification would be helpful.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Manuscript Revisions

We thank all four reviewers for their constructive feedback and for recognizing the value of MetaFauna's structured observation metadata [v9Uv, 13T8, w2fd, 2LJW] and controlled experimental design [v9Uv, w2fd, 2LJW]. During the discussion period, we uploaded the complete dataset annotation package, reported quantitative annotation and split audits, conducted additional experiments, and updated the manuscript accordingly.

1. Dataset Availability and Reproducibility (13T8 Q4, Q7)

We uploaded the MetaFauna-v1-annotations.zip package to the Supplementary Material, containing identity labels, structured metadata, source-image mappings, and fixed train/query/gallery splits for all images in MetaFauna. Our anonymized repository provides a single-command tool to download the original images from LILA and reconstruct the animal crops. This allows reviewers to inspect the annotations and reconstruct the dataset.

2. Annotation Reliability and Split Verification

We provided quantitative checks for the identity and metadata annotations:

  • Identity annotations (2LJW Q1; 13T8 Q2): Initial agreement reached 91.0% (=0.825) across 30,217 candidate pairs. Independent audits found zero incorrect merges across 270 identity clusters and zero train/test identity conflicts across 5,781 retrieved candidate pairs, with two missed test-set associations resolved prior to finalizing the benchmark.
  • Metadata annotations (13T8 Q2; w2fd W2): Independent annotators achieved 95.8%–98.1% agreement across G1 fields. G2 and G3 values were checked through independent computation and source verification.
  • Evaluation splits (2LJW Q2; 13T8 Q3; w2fd W2): We examined all 9,935 Hyena and Leopard images and confirmed 655 near-duplicate or same-encounter groups, none of which span query and gallery. For Snapshot Safari, we also reported the number of identities with cross-site and temporally separated matches.

3. Additional Experiments Addressing Reviewer Questions

ExperimentReviewer / ConcernEvaluation and Key Outcome
Backbone scaling & pretrainingv9Uv (W2); w2fd (Q3, W3)DINOv3 (ViT-L to ViT-7B) and SigLIP 2; MetaRouter retains +2.55 to +4.12 mAP over Shared LoRA
External datasetsw2fd (Q1, W4); 13T8 (Q6)MetaWild (6 species) and CzechLynx; highest mAP on all 7 species (+3.84 mean mAP over Shared LoRA)
Predicted & corrupted metadataw2fd (Q2); 13T8 (Q5)Automated G1 prediction retains 66.93% mAP (+2.63 over Shared LoRA); stable under moderate errors
Foreground-only evaluation13T8 (Q3); 2LJW (Q2)Hyena and Leopard with backgrounds removed; +5.9 and +4.3 mAP over Shuffled MetaRouter
Cross-site & temporal retrieval2LJW (follow-up)Snapshot Safari (24h, 7d, different sites); +2.7 to +3.2 mAP over Shared LoRA

4. Manuscript Revisions We moved the individual metadata-field analysis from the Appendix into the main paper (Fig. 5, per v9Uv Q2), added the backbone-scaling summary (Table 4), clarified visual-only inference (p=1) and MetaRouter's role as a reference framework alongside MetaFauna as the primary dataset contribution, and documented all audits and additional experiments in Appendices A, C, and F.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.