MegaDescriptor: Wildlife Re-ID Foundation

• MegaDescriptor is a Swin-transformer based model for wildlife re-identification that uses deep embeddings and cosine similarity to identify individual animals across species.
• It leverages metric learning with ArcFace loss on 29 datasets, establishing a robust global descriptor that consistently outperforms generic pre-trained models.
• The model serves as a foundational backbone that can be fused with local matchers, enabling enhanced open-set recognition and modular adaptation in multispecies contexts.

MegaDescriptor is a foundation descriptor for wildlife re-identification: a Swin-transformer-based model trained specifically for individual animal identification across many species rather than for generic visual semantics. It was introduced in the WildlifeDatasets toolkit as “the first-ever foundation model for individual re-identification within a wide range of species,” with the practical aim of replacing fragmented species-specific pipelines by a reusable embedding model that supports zero-shot retrieval, benchmarking, and downstream adaptation (Čermák et al., 2023). Subsequent work has treated MegaDescriptor as a strong domain-specific baseline, a frozen backbone for open-set recognition, and a global similarity component within calibrated fusion systems, while also using it as the principal reference point for newer multispecies models (Cermak et al., 2024, Miyaguchi et al., 15 Sep 2025, Otarashvili et al., 2024).

1. Origins and intended scope

MegaDescriptor emerged from a specific methodological gap in wildlife computer vision. Animal re-identification had been dominated by species-specific models, inconsistent evaluation protocols, and workflows that imposed repeated costs for data collection, curation, training, deployment, and maintenance. WildlifeDatasets framed MegaDescriptor as a standardized, transferable descriptor meant to work across a wide range of species and to be usable either directly in retrieval settings or as a pre-trained initialization for downstream re-ID tasks (Čermák et al., 2023).

The original benchmark context is unusually broad. WildlifeDatasets provides access to 31 publicly available wildlife datasets as of the end of September 2023, while the core MegaDescriptor experiments train and evaluate on 29 datasets, excluding Drosophila and SeaTurtleID2022 from some ablations. The toolkit supports closed-set, open-set, and disjoint-set splits, as well as time-aware splits when timestamps are available. This broader infrastructure is important because MegaDescriptor was not introduced as an isolated model, but as part of an attempt to regularize wildlife re-ID experimentation and deployment (Čermák et al., 2023).

Later papers preserve this framing but sharpen the terminology. WildFusion describes MegaDescriptor as a Swin-based foundational model for animal re-identification, originally trained on over 30 datasets with ArcFace loss, and uses MegaDescriptor-L-384 as a global/deep similarity model rather than as a local matcher (Cermak et al., 2024). DS@GT AnimalCLEF correspondingly treats it as a domain-specific foundation model for wildlife re-identification, contrasting it with the general-purpose DINOv2 manifold (Miyaguchi et al., 15 Sep 2025).

2. Architecture, embedding formulation, and training recipe

MegaDescriptor is architecturally centered on Swin Transformer backbones and metric learning. WildlifeDatasets reports a model-selection process comparing Swin-B and EfficientNet-B3, with ArcFace and transformer-based backbones outperforming Triplet loss and the CNN baseline. The operational design is deliberately simple at inference time: wildlife re-identification is formulated as embedding learning followed by nearest-neighbor retrieval in embedding space, and the matching strategy is described as “essentially” a 1-nearest-neighbor classifier using cosine similarity (Čermák et al., 2023).

In later reuse, this cosine-similarity view is made explicit. WildFusion states that MegaDescriptor and DINOv2 provide deep embeddings whose pairwise cosine similarity serves as the global similarity score:

$$s_G(x_0, x_1) = \frac{f(x_0) \cdot f(x_1)}{\|f(x_0)\| \|f(x_1)\|}.$$

This clarifies how MegaDescriptor is operationalized as a descriptor model: it produces a global embedding, and matching is performed through similarity in that embedding space rather than through explicit local correspondence search (Cermak et al., 2024).

The original training regimen in WildlifeDatasets uses the union of 29 publicly available datasets, each split 80/20 into reference and query sets while preserving a closed-set setting. Optimization uses SGD with momentum $0.9$, 100 epochs, cosine annealing learning-rate scheduling, and batch size 128. Model selection is extensive: the paper reports a grid over backbone $\{\texttt{Swin-B, EfficientNet-B3}\}$, learning rate $\{\texttt{0.01, 0.001}\}$, ArcFace margin $\{\texttt{0.25, 0.5, 0.75}\}$, ArcFace scale $\{\texttt{32, 64, 128}\}$, Triplet mining $\{\texttt{all, semi, hard}\}$, and Triplet margin $\{\texttt{0.1, 0.2, 0.3}\}$, totaling 72 settings per dataset and 2088 training runs. The best ArcFace setting is reported as $lr=0.001$, $m=0.5$, and $0.9$0 (Čermák et al., 2023).

WildlifeDatasets distributes multiple MegaDescriptor flavors through HuggingFace. The paper-level flavor ablation analyzes Tiny, Small, Base, Large-224, and Large-384 variants, and concludes that both model size and input resolution generally improve performance, with MegaDescriptor-L-384 the best overall configuration (Čermák et al., 2023). In AnimalCLEF, MegaDescriptor (Large) is described more concretely as Swin-L with approximately 229 million parameters, output dimension 1024, and training on 29+ public animal re-identification datasets; the model actually used there is MegaDescriptor-L-384 (Miyaguchi et al., 15 Sep 2025).

3. Benchmark performance in WildlifeDatasets

Within WildlifeDatasets, MegaDescriptor is positioned against local descriptors, deep metric-learning baselines, and generic pre-trained visual models. The local baselines are SIFT and SuperPoint; the deep baselines are ArcFace and Triplet loss; the foundation-style comparisons include ImageNet-1k/Swin-B, CLIP, and DINOv2. The central conclusion is categorical: MegaDescriptor-L “outperforms all methods on all 29 datasets” in the zero-shot retrieval setting used in the paper (Čermák et al., 2023).

The comparative margins against DINOv2 are especially large on many datasets. WildlifeDatasets reports, for example, 66.48 vs 11.20 on BelugaID, 97.82 vs 52.75 on BirdIndividualID, 91.10 vs 50.38 on CTai, 77.81 vs 3.23 on HumpbackWhaleID, 91.18 vs 33.57 on SeaTurtleIDHeads, 78.66 vs 34.05 on SealID, and 74.40 vs 12.26 on ZindiTurtleRecall. These results underpin the later characterization of MegaDescriptor as a domain-specific embedding space rather than merely a larger backbone (Čermák et al., 2023).

The paper also provides many dataset-specific MegaDescriptor-L scores, including 99.93 on AAUZebraFish, 94.33 on ATRW, 99.54 on Cows2021, 83.17 on GiraffeZebraID, 78.41 on HyenaID2022, 75.58 on LeopardID2022, 67.42 on NDD20, 62.02 on WhaleSharkID, and 100.0 on both OpenCows2020 and SMALST. At the same time, the distribution of results shows that the descriptor does not collapse task difficulty: harder datasets remain difficult, including HappyWhale at 34.30, NOAARightWhale at 40.26, and NyalaData at 36.45 (Čermák et al., 2023).

Model scaling is also systematic rather than anecdotal. The appendix reports that larger MegaDescriptor variants usually improve performance, with only minor underperformance of larger models on FriesianCattle2017, LeopardID2022, and MacaqueFaces by 2.53%, 0.48%, and 0.08%, respectively. Resolution is a substantial factor: on BelugaID, the progression runs from 33.12 (T-224) to 66.48 (L-384); on HumpbackWhaleID from 28.12 to 77.81; and on ZindiTurtleRecall from 26.77 to 74.40. WildlifeDatasets therefore presents MegaDescriptor-L-384 as the canonical high-capacity form of the model (Čermák et al., 2023).

4. MegaDescriptor as a global similarity model and open-set backbone

A major development after the original release is the reuse of MegaDescriptor as a modular component rather than only as a standalone retriever. WildFusion is explicit on this point: MegaDescriptor is treated as a global/deep similarity model, not as a local matcher. Its role is to produce deep embeddings whose cosine similarity can be fused with local-matching similarity from LightGlue+ALIKED, LightGlue+DISK, LightGlue+SuperPoint, and LoFTR. The fusion mechanism is score-level rather than feature-level:

$0.9$1

with equal weights in the baseline and calibration performed by isotonic regression with PCHIP interpolation; logistic regression is also evaluated (Cermak et al., 2024).

This calibrated fusion substantially changes the empirical position of MegaDescriptor. WildFusion reports MegaDescriptor-L-384 at 75.50% average accuracy, WildFusion using only local matching scores at 78.48%, and WildFusion using all scores at 83.99%, corresponding to a gain of 8.49 percentage points over MegaDescriptor-L-384. The full system outperforms MegaDescriptor on 16 of 17 datasets, with reported gains of +25.19 on NDD20, +18.28 on WhaleSharkID, +17.29 on SeaStarReID2023, +14.35 on SealID, +13.11 on LeopardID2022, +12.77 on GiraffeZebraID, and +12.06 on HyenaID2022; the cited exception is ZindiTurtleRecall, where MegaDescriptor-L scores 74.24 and WildFusion scores 71.90 (Cermak et al., 2024).

WildFusion’s ablation results further indicate that MegaDescriptor and local matchers are complementary rather than redundant. The paper reports 75.5 for MegaDescriptor-L only, 83.0 for MegaDescriptor-L + LG-ALIKED, 81.4 for MegaDescriptor-L + LoFTR, and 84.0 for MegaDescriptor-L + all local methods. In the paper’s interpretation, MegaDescriptor remains a strong global embedding but can miss fine-grained local evidence that local matchers preserve; calibration makes those heterogeneous signals comparable and therefore fuseable (Cermak et al., 2024).

In open-set evaluation, DS@GT AnimalCLEF places MegaDescriptor in a different role: a frozen backbone encoder whose CLS embedding is fed into a K-nearest-neighbor classifier with robust thresholding. The official competition score is

$0.9$2

where BAKS measures balanced accuracy on known samples and BAUS on unknown samples. In this pipeline, MegaDescriptor baseline scores 0.28528 on the public leaderboard and 0.25967 on the private leaderboard, exceeding the DINO baseline at 0.24371 public and 0.18856 private. A triplet-learning projection head improves the specialized MegaDescriptor model by 0.13 points, while DINOv2 receives only a 0.03 improvement on averaged BAKS and BAUS according to the authors’ discussion (Miyaguchi et al., 15 Sep 2025).

DS@GT interprets these results as evidence that post-hoc metric learning is highly contingent on the initial quality and domain-specificity of the backbone manifold. MegaDescriptor is described as easier to refine: it yields lower training and validation losses, reduces the number of mined semi-hard triplets more strongly over training, and provides a manifold already containing more clustering structure before triplet learning. A plausible implication is that MegaDescriptor’s wildlife-specific pretraining makes its embedding geometry more amenable to small-data refinement than general-purpose self-supervised features (Miyaguchi et al., 15 Sep 2025).

5. Relationship to later multispecies re-ID models

Later multispecies work treats MegaDescriptor as the closest prior foundation-style wildlife re-ID baseline, but no longer as the endpoint. The clearest example is MiewID, which trains a single embedding network on 49 species, 37,138 individual animals, and 225,374 images using an EfficientNetV2-M backbone with sub-center ArcFace and dynamic margins. In that paper’s framing, MegaDescriptor is the recent descriptor model intended to generalize across species, including unseen ones, and thus the relevant target for zero-shot comparison on held-out species (Otarashvili et al., 2024).

The reported comparison is decisive within that study’s protocol. For 33 unseen species that MegaDescriptor had not been trained on, MiewID trained without the held-out species outperforms MegaDescriptor-L-384 on all 33 species, with an average top-1 improvement of 19.2% per species. Examples include bottlenose dolphin from 17.7 to 71.5, dusky dolphin from 28.5 to 84.8, hyperoodon ampullatus from 35.3 to 72.9, japanese monkey from 50.8 to 84.6, melon-headed whale from 38.2 to 92.1, spinner dolphin from 19.1 to 95.2, and white shark + fin dorsal from 11.2 to 64.9 (Otarashvili et al., 2024).

This does not negate MegaDescriptor’s earlier status; rather, it clarifies its historical position. In MiewID’s account, MegaDescriptor represents the prior state of the art for generalized animal re-ID descriptors, especially on unseen species. MiewID’s gains are attributed to a much larger community-curated multispecies corpus and a different architectural and training design. The paper also reports that its multispecies model achieves an average gain of 12.5% in top-1 accuracy over species-specific models, suggesting that the field’s center of gravity had begun to move from generalized descriptor transfer alone toward large-scale multispecies training (Otarashvili et al., 2024).

6. Limitations, evaluation caveats, and broader reuse

The most important caveat in the original MegaDescriptor paper is evaluation scope. WildlifeDatasets explicitly does not claim that MegaDescriptor solves open-set re-identification; its main evaluation setting is closed-set retrieval in which all query identities appear in the reference set. The paper also notes that some datasets are already effectively solved or close to solved, which weakens their future benchmarking value, and it emphasizes the sensitivity of metric learning to hyperparameters: some ablation settings collapse to mean accuracies of 6.4%, 6.1%, and 4.0% (Čermák et al., 2023).

A second limitation is that strong performance as a descriptor does not imply universal dominance in all zero-shot wildlife organization tasks. In the Animal Detect study on zero-shot wildlife sorting, MegaDescriptor-L-384 is one of three backbones evaluated alongside CLIP ViT-L/14 and DINOv2 ViT-G/14 for clustering and continuous similarity ordering. However, the best reported clustering result in the provided text is obtained by DINOv2 + UMAP + GMM, which reaches 443/500 correctly grouped images, 0.886 accuracy, and 0.874 macro-F1; the excerpt does not report a separate headline metric for MegaDescriptor. This suggests that MegaDescriptor is a domain-relevant backbone for zero-shot wildlife embedding, but not necessarily the top-performing representation under every downstream unsupervised organization pipeline (Markoff et al., 16 Oct 2025).

A third limitation is conceptual rather than empirical: later work increasingly separates global descriptor quality from local geometric evidence. WildFusion’s results show that MegaDescriptor benefits substantially from calibrated fusion with local matchers, and its zero-shot protocol excludes MegaDescriptor entirely because MegaDescriptor had already been trained on the relevant data. In that setting, local-only WildFusion achieves 76.2% mean accuracy across 17 datasets, 0.7 points above MegaDescriptor-L-384 according to the text, and only 2.3 points below dataset-specific calibration. This suggests that the field no longer treats global embeddings alone as sufficient for all animal re-ID regimes, especially when fine-grained local markings dominate identity evidence (Cermak et al., 2024).

Taken together, these later results define MegaDescriptor’s mature significance. It remains a foundational wildlife-specific descriptor and a strong domain-specific embedding manifold; it established a benchmark substantially above CLIP, ImageNet, and DINOv2 in the 2023 WildlifeDatasets setting; and it continues to function as a baseline, backbone, and fusion component in subsequent work. At the same time, later research indicates two distinct directions beyond it: calibrated combination with local matchers for fine-grained evidence aggregation, and larger multispecies training corpora that improve transfer to unseen species (Čermák et al., 2023, Cermak et al., 2024, Otarashvili et al., 2024).