This paper presents a three‑stage meta‑algorithm that addresses open‑set individual animal re-identification. The cascade first employs WildFusion to fuse calibrated global-local similarity scores, then feeds concatenated MegaDescriptor‑L and MIEW embeddings into an XGBoost classifier, and finally refines predictions with species‑specific Dual‑Backbone models fine‑tuned using an ArcFace angular‑margin loss. On the AnimalCLEF 2025 challenge, which includes loggerhead sea turtles, fire salamanders, and Eurasian lynxes and exhibits a pronounced long‑tail imbalance, the proposed method achieved a private score of 67.42% and a public score of 65.11%, ranking 2nd out of 172 teams. Ablation analysis shows cumulative improvements of +21 percentage points (pp) from WildFusion over a MegaDescriptor baseline, +2.4 pp from XGBoost, and +3 pp from the Dual‑backbone ArcFace stage. These results demonstrate that species‑aware stacking of heterogeneous cues (global descriptors, calibrated local matches, tabular neighbor context, and metric fine‑tuning) yields a robust and scalable solution for non‑invasive wildlife monitoring.
Individual animal re-identification (Animal Re-ID) is the task of recognizing spiencdiificviduals in images. Accurate identity assignment is critical to ecology and wildlife conservation because it enables monitoring of population size, migration routes, and behavioral patterns of rare specie1s].iInnsitu [ Human Re‑ID universal biometric cues such as faces or fingerprints are available, whereas these markers are not directly applicable to animals. Instead, recognition relies on unique natural markings—spot and stripe patterns, carapace mosaics, and similar traits—that vary markedly with viewpoint, pose, and illumination2[]. The problem is compounded by a shortage of labeled data: collecting and annotating photographs of individual animals is labor‑intensive, so Animal Re-ID datasets are several orders of magnitude smaller than those used in Person Re‑ID [3].
Traditional biological approaches—ringing, tagging, and DNA analysis—are reliable but invasive and unsuitable for large‑scale monitor4in].gE[arly computer‑vision algorithms addressed one species at a time and relied on handcrafted features, which does not scale. With deep neural networks (first CNNs, later Vision Transformer5s])[Human Re‑ID achieved a high level of accuracy, yet direct transfer to animals proved inefective: the class set is open, inter‑individual diferences are subtle, and intra‑species variability is high [6]. These factors motivated specialized methods for Animal Re-ID.
The AnimalCLEF 2025 competition [78,] poses a multi‑species challenge: identifying loggerhead sea turtleCsaretta caretta (Greece), fire salamandersSalamandra salamandra (Czech Republic), and Eurasian lynxeLsynx lynx (Czech Republic) [9]. For each input image the system must decide whether the depicted animal belongs to a knoinwdnividual in the training database (tdahteabase set or ”gallery”) or represents naew individual; if known, the correct ID must be returned. Consequently, the task combines classical Re‑ID with aonpen‑set component. Performance is evaluatedBbAyKS (balanced accuracy on known samples) anBdAUS (balanced accuracy on unknown samples); the final score is the geometric mean of these two metrics [7].
The meta‑algorithm proposed in this study—combining WildFus1io],nX[GBoost, and a Dual‑backbone model with an ArcFac1e0[] head—ranked second among 172 teams, achievinpgraivate score of 67.42% and apublic score of 65.11% (team Webmaking). Subsequent sections describe the employed methods in detail (Sec3.) and present a step‑by‑step analysis of the contribution made by each component to the final performance (Sec4.).
The primary goal of AnimalCLEF 2025 is to advance automated biodiversity monitoring, in particular the tracking of individual animals captured by camera traps and other imaging devices. Precise identification of individuals is pivotal in ecology: it enables reliable estimates of population size, migration routes, and behavioral profiles that underpin both scientific studies and conservation measures. Existing algorithms, however, tend to overfit to background or illumination cues and lose accuracy when applied to novel conditions. Consequently, the competition focuseusnoivnersal Re‑ID approaches that can generalize across habitats and reliably recognize animals in a wide range of environments. Participants could either rely solely on the limited competition data or improve their models by pre‑training on the large external datasetWildlifeReID‑10k [11]. Overall, AnimalCLEF 2025 serves as a benchmark for state‑of‑the‑art computer‑vision methods and continues the LifeCL8E]Fse[ries that expands the role of AI in wildlife monitoring.
For training and pre‑training, participants were provided with the large WildlifeREID‑10k dataset containing roughly 140,000 images of more than 10,000 individual animals across many sp1e1c]i.es [ This external resource can be regarded as an extended training set. The competition data were collected specifically for AnimalCLEF 2025 and split into two parGtasl:lery, which holds annotated images of known individuals and simultaneously serves as the training set and the gallery for matching, and Query (Tab. 1).
LynxID2025. This subset comprises 2957 training images of Eurasian lynx and 946 query images (3903 in total). The training split covers 77 unique individuals, with an average of 38 photographs per individual; the distribution is unbalanced—some animals have only a single image, whereas one individual appears in 353 shots. Image orientation is recorded for every picturrigeh(tl,efft,ront, back, orunknown). Capture dates are not provided (dtahtee field is empty).
SalamanderID2025. This subset contains 1388 training images of fire salamanders and 689 query images (2077 in total). The training split includes 587 unique individuals; the average is ~2.4 images per individual, the median is 1, and the maximum is 12. Orientation latbope,lsbo(ttom, left, right) are available for all images. Capture dates span 2017–2023 in the training set and extend to 2024 in the query set, enabling temporal analysis of the data collection period (for example, training pictures cover 2017–2023, while query images include shots made up to the end of 2024).
SeaTurtleID2022. This is the largest subset: 8729 training photographs of loggerhead sea turtles and 500 query images (9229 in total). The training split represents 438 unique sea turtles (mean 19.9 images per individual; median 13). Orientation labels inclelft,udreight, front, top, and composite directions such astopleft ortopright; orientation is missing for all 500 query images. Capture dates are present for almost every photo, ranging from 2010 to 2024, reflecting the long‑term nature of data collection.
The three subsets difer markedly. Lynx ofers fewer individuals but more images per individual, whereas Salamander provides many individuals yet mostly single‑image observations. Sea Turtle occupies an intermediate position in terms of individual count, but its total image volume is the largest. Such heterogeneity in size, orientation metadata, and temporal coverage underscores the need for adaptive identification strategies tailored to each species.
AnimalCLEF 2025 employs two metrics that jointly assess recognition quality on known and new individuals. BAKS is the per‑class balanced accuracy over query images whose individuals are present in the gallery. BAUS is the balanced accuracy over query images belonging to new individuals absent from the gallery. The final ranking score is the geometric mean of BAKS and BAUS. The organizer split the query set into an open (pub)lpicortion comprising about 31% of the images and a hiddperniva(te) portion comprising the remaining 69%. Only the private leaderboard determined the final standings, preventing overfitting to the public subset.
The task of individual animal identification had been explored before AnimalCLEF 2025. A notable predecessor is the Happywhale—Whale & DolphinID competition (Kaggle 2022), which required distinguishing thousands of individuals from 24 marine‑mammal species using the mAP metric; the task sufered from a strong class imbalance but lacked an open‑set component [12].
Between 2022 and 2024 several species‑specific re‑ID datasets were released together with mini‑competitions, including Leopard ID and Hyena ID from WildMe & LILA Scie1n3c,e14[] and SeaTurtleID [15]. SeaTurtleID first introduced time-aware closed- and open-set splits later adopted by AnimalCLEF. An internal benchmark demonstrated 86.8% closed-set accuracy when using a Hybrid Task Cascade equipped with an ArcFace encoder, highlighting the challenges of long-term individual tracking even within a single species [15].
Modern Animal Re-ID methods rely on deep networks that extract image embeddings, i.e., compact feature vectors unique to e aincdhividual. Two principal categories of such features exgislotb:al descriptors that summarize the entire image alnodcal matches that align distinctive regions.
A prominent global approach MisegaDescriptor [16]. This supervised model is trained on a collection of many datasets (>10k individuals, ~140k images). Its backbone is a Swin‑L Transformer with384 × 384 input and about 229 M parameters. Essentially a Vision Transformer tuned for Animal Re-ID, it markedly outperforms generic models such as CLIP and DINOv2 [16].
An alternative global encodeMrIiEsW (Multi‑species Individual Embeddings Wild, MiewID‑msv3). This compact EficientNet‑V2 [17] CNN (about 51 M parameters) is trained with a contrastive Sub‑center ArcFace loss on a dedicated dataset of 64 species (225k photos, 37k individuals). Unlike MegaDescriptor, which is trained per species, MIEW is optimized as a single multi‑species model. Experiments show that this unified model surpasses species‑specific training by an average of 12.5% top‑1 and, more importantly, generalizes better to unseen species: on unknown taxa MIEW outperforms MegaDescriptor by 19.2% top‑1 accuracy [3], demonstrating its ability to capture universal cues useful for any animal.
Global descriptors have limitations: they may miss fine‑grained individual patterns. To compensate, local methods match image regions that carry unique markings. The modWerinldFusion framework combines global and local information eficient1l]y. [It fuses (i) cosine similarity of global embeddings (e.g., MegaDescriptor or DINOv2) and (ii) local keypoint correspondences obtained with matchers such as LoFTR [18] or LightGlue1[9]. After isotonic calibration, the two similarity sources are merged into a single score. In a zero‑shot setting WildFusion exceeds the pretrained MegaDescriptor‑L, confirming that hybrid cues can substantially improve Re‑ID performance [1].
A common path to higher accuracy is ensemble learning. In this work several ways of combining
embeddings from MegaDescriptor and MIEW were explored. The best result was achieved by a
meta‑algorithm that blends predictions from (
MegaDescriptor‑L‑384 is a foundation model for Animal Re-ID introduced16i]na[nd released on Hugging Face [20]. The backbone isswin_large_patch4_window12_384 with384×384 px input and 228.8 M parameters; it outputs a 1536‑dimensional L2‑normalized embedding suitable for cosine comparison.
The network was trained in a supervised manner with an ArcFace‑style margin loss on the aggregated WildlifeDatasets corpus comprising 29 public datasets (~140k images, >10k individuals, 23 species). Merging such diverse sources exposes the model to wide variations in viewpoint, illumination, and marking patterns, thereby improving embedding generality. The authors report that MegaDescriptor‑L‑384 consistently outperforms CLIP (ViT‑L/336) and DINOv2 (ViT‑L/518) on all 29 benchmarks [16].
In practice, deployment requires only standard preprocessing: re3s8iz4e×3t8o4, convert to a tensor, and normalize to mean(s0.485, 0.456, 0.406) and standard deviatio(n0s.229, 0.224, 0.225). A single forward pass then produces the embeddin20g][. The CC‑BY‑NC‑4.0 license permits non‑commercial use and modification, making MegaDescriptor‑L‑384 a strong out‑of‑the‑box global descriptor within the pipeline.
WildFusion [16] addresses a core limitation of purely global embeddings in Animal Re-ID—their sensitivity to background and illumination. By combining global image similarity with precise local keypoint verification, the method sharply reduces false matches between difeinrednivtiduals while recovering true correspondences under strong viewpoint changes. A detailed analysis of its impact on the final private score is presented in Sec4..
The algorithm comprises two stages. First, a fast cosine search in the MegaDescriptor‑L‑384 embedding space (Sec.3.2) selects the= 300 most similar gallery images (candidates). Each “query / candidate” pair is then evaluated by five independent local pipeLlionFeTsR:, SuperPoint [21], ALIKED [22], DISK [23], andSIFT [24]. For LoFTR, images are converted 1t9o2×192 px grayscale, whereas the other pipelines operate 5o1n2×512 px color inputs. The local scores and the normalized global similarity undergo isotonic calibration and are subsequently fused linearly into a single probabilistic score.
An open‑source implementation is provided in twhieldlife‑tools package [25] in the wildlife‑datasets repository1[6]. Released under the GPL‑3.0 license, the code requires no additional training, making WildFusion easy to integrate into an existing pipeline.
On top of WildFusion probabilitikes-r,eciprocal re-ranking [26] is applied. Three similarity matrices are available: “quer×ygallery”, “galler×yquery”, and “galler×ygallery”. For each image a list of the 1 = 20 most similar neighbors is formed according .toA gallery imag e retains only those neighbors that simultaneously pl acweithin their first 2 = 6 positions; the resulting mutual set is denotedrc() . An analogous procedure yielrdcs() for every query, as the symmetric “gall×erqyuery” matrix enables reciprocity checks in the reverse direction.
The final similarity between a quer yand a gallery imag eis calculated as
This linear convex combination suppresses incidental matches caused by pose, masking, or background, while requiring no additional model training.
Score‑CAM [27] heatmaps in Fig.1 indicate that MegaDescriptor‑L‑384 focuses on compact texture regions, whereas MIEW‑msv3 distributes attention across fine‑grained spots and extended contours. The theoretical complementarity of these spatial patterns motivates a direct concatenation of the two embeddings (ℝ3688), with each component pre‑normalized by it2snorm [3, 16].
For every image, the feature vector includes (i) the concatenated global embe d=d1i0ng,c(oiis)ine distances to the ten nearest gallery images together with the corrensepiognhbdoinrgidentifiers passed as categorical features1, and (iii) one‑hot representations of view orientation and dataset membership (lynx, salamander, sea turtle). The key idea is that gradient boosting can non‑linearly merge global descriptors, local density information in the embedding space, and categorical data on thinedcivloidsueaslts. The maximum depth was capped at 6, which prevents the model from memorizing category values via long split chains.
Incorporating both global descriptors enhances feature diversity: the transformer‑based MegaDescriptor captures coarse texture patterns, whereas the CNN backbone of MIEW remains sensitive to 1The columnsnn_id_1…10 are cast tpoandas.Categorical and processed by XGBooste’snable_categorical option, which learns optimal subset splits instead of numerical thresholds. 0 Baseline provided by the competition organizers: 30.90
MegaDescriptor‑L threshold 0.6 for all 1 Baseline: MegaDescriptor‑L cosine nearest neighbor with 40.59
per‑species thresholds 2 WildFusion global + local similarity fusion with thresholds 61.72 3 k‑reciprocal re‑ranking (Lynx only) applied to WildFusion 62.09 4 XGBoost meta‑classifier on MegaDescriptor + MIEW embed- 64.44
dings; WildFusion confidence adjustment 5 Dataset-specific meta‑algorithm that combines WildFusion, 67.42
XGBoost and Dual-backbone ArcFace — +9.69 point‑wise diferences [3, 16]. XGBoost trained on this concatenation, augmented with local density features and metadata, produces a consistent improvementpirnivtahteescore andpublic score relative to a single embedding baseline and to WildFusion alone4()S. ec.
The two previous ensemble components (WildFusion, S3e.c3., and XGBoost, Sec3..5) rely onfixed embeddings obtained without species‑specific fine‑tuning. Although this delivers high baseline accuracy, the capacity to adapt to species‑specific visual patterns remains limited. The Dual‑backbone model addresses the opposite need: it refines featupreersspecies and thus complements the rigid matching scheme of WildFusion and the tabular classifier XGBoost. Methodologically the model unites deep metric optimization via ArcFac1e0][ with the direct feature focus provided by two heterogeneous backbones.
The first stream employsMegaDescriptor‑L‑384, reliable at capturing global textures; the second employs MIEW‑msv3, sensitive to fine pointwise details. Both classification heads are removed, and their outputs after individual BatchNorm layers are concatenated into a 3688‑dimensional vector.
A compact ArcFace head is placed on top of the joint space. ArcFace maximises inter-class angular margins in the embedding space, imposing a strict separability criterion. This margin‑based approach is particularly efective under the small‑sample conditions characteristic of AnimalCLEF 2025 [10].
WildFusion depends on a calibrated “global + local” heuristic, and XGBoost on tabular aggregation of fixed embeddings and metadata. Three independent Dual-backbone models, one trained for each species, supply descriptors tailored to their respective AnimalCLEF 2025 subsets and thereby improve the robustness of the ensemble.
BVRA/MegaDescriptor‑L‑384; batch 32, input384×384) was evaluated with both shared and thresholds (Sec.3.2). Each gallery image =( 13,074 ) and each query image (= 2135 ) was encoded as a 1536‑dimensional vector. Cosine similarity was computed between every query vector and every gallery vector; the most similar gallery image provided the candidate identity for the query. If the similarity fell below the threshold, the lnaebwe_lindividual was assigned.
A grid search over a single global threshold in the range 59.0–74.5% yielded the best result at 74.0% (public 35.76%, private 37.32%). Separate thresholds were then explored for each taxon: Lynx 50–90%, Salamander 60–90%, Sea turtle 74–90%. The optimal triplet (Lynx 65.5%, Salamander 77.0%, Sea turtle 74.5%) produced a public score of 37.92% and a private score of 40.59%, establishinbagstehliene for subsequent steps.
Preliminary analysis of Fig2.revealed that submissions cluster vertically: runs withfsraimc-ilar tions of new_individual predictions tend to yield comparable public scores, and the highest–scoring points concentrate around species‑specific share7s0o%f (lynx),60% (salamander) an6d0% (sea turtle). Therefore, at all later steps (including the WildFusion, XGBoost confidence gate, and the final cascade) thresholds were selected so as to preserve these empirically favorable ratios. This policy explains the multiple vertical stripes visible in 2F:ige.ach stripe marks a family of submissions that intentionally maintain the samenew_individual quota while refining other components of the pipeline.
Possible future work includes replacing manual threshold search with probability–calibration techniques such asPlatt scaling ortemperature scaling, which learn a monotone mapping on the validation split and may further stabilize the species‑spenceificw_individual ratios without exhaustive grid search.
[16, 25] is employed at this step (Sec3.3).
First, MegaDescriptor‑L‑384 retrieves t he= 300 gallery candidates with the highest cosine similarity; all 15,209 images (13,074 gallery + 2135 query) are encoded as 1536‑dimensional vectors. Each ”query / candidate” pair is then evaluated by five independent local matchers: SuperPoint–LightGlue, ALIKE–LightGlue, DISK–LightGlue, SIFT–LightGlue (all at×551122 px RGB) and LoFTR (192× 192 px grayscale). Figur3e shows that the detectors yield a comparatively small overlap of correspondences; this diversity underlies the gain obtained after isotonic calibration and score fusion. Calibrating the five matchers on the validation split required 3h on a single A100 GPU; tqhueerfyul×lgallery evaluation took a further 27h.
After combining global and local signals, thresholds fonretwh_eindividual label were tuned separately per species. The optimal values were Lynx 39.5%, Salamander 12.0%, and Sea turtle 16.0%.
This configuration achieved a58.98% public score and 6a1.72% private score, yielding a +21 pp improvement over the baseline in Ta2b..Notably, the tuned WildFusion stage alone would already have secured an 8th place finish on the final leaderboard, even before adding the later cascade components.
The results confirm that merging global embeddings with complementary local keypoints is crucial for the substantial performance gain observed on AnimalCLEF 2025.
matrix was computed for the Lynx subset and re‑ranking was applied according to the scheme of Z. Zhong [26]. The method parameters were fixed to the first neighborhood rad iu1s= 20, the reciprocity radius 2 = 6, and the Jaccard weigh=t 0.1 in the linear combination with the original WildFusion score (Sec.3.4).
Lynx was selected because its images share an artificially uniform black background, which increases the risk of pose‑driven false matches; reciprocal filtering helps to attenuate this artifact. Building the gallery × gallery matrix required an additional 38 GPU‑hours, so the procedure was not executed for Salamander or Sea turtle.
The gain, although modest, was positive: the public score rose from 58.98% to 59.09% (+0.11 pp), and the private score from 61.72% to 62.09% (+0.37 pp). This confirms the value of mutual neighbor filtering, yet the improvement did not justify the computational cost; all subsequent steps therefore relied on the original WildFusion scores (for Salamander or Sea turtle).
image a dense feature vector3o7f18 dimensions was assembled: the 3688‑D concatenation of MegaDescriptor‑L‑384 and MIEW‑msv3 embeddings, 10 cosine distances to the nearest gallery images, 10 categorial identifiers of those neighbors, and 10 one‑hot categoroireiesn(t7ation flags + 3 dataset flags). A unified probability scale simplifies the tuning of the cascade.
An XGBoost model was trained with depth =6, 0.15 , = 2.0 , tree_method=gpu_hist and the multi:softprob objective; the best iteration was reached at round 296. Validation followed a “single image per individual” split (Sec.3.5).
During inference the posterior probabilities of XGBoost acted as a confidence gate on top of WildFusion. Species‑specific thresholds were tuned empirically: for Salamander, the WildFusion label was replaced when XGBoost confidence exceeded 20%; for Sea turtle, when it exceeded 95%. This cascaded refinement raised thepublic score to 61.89% and thperivate score to 64.44%, adding +2.80 pp and +2.35 pp, respectively, over the pure (re‑ranked) WildFusion.
rithm For each taxon an individual Dual‑backbone network was trained that combines MegaDescriptor‑L‑384 with MIEW‑msv3. After separateBatchNorm layers the two vectors were concatenated into a 3688‑dimensional feature, which was fed to a compAarcctFace head (Sec.3.6). The head parameters were fixed per dataset(:, ) = (64, 0.5) for Lynx and Sea turt(l3e0,, 0.35) for Salamander.
Augmentation pipelines were tailored to the visual specifics of each daLtyansxe:tb.ackground‑mask removal followed bRyandomResizedCrop with scale≥ 0.9. Salamander: rotation according to the orientation field and moderate cropping that preserves key anatomical regSieoantsu. rtle: moderate cropping plus horizontal flip. All datasets additionally recCeoilvoerdJitter andCoarseDropout (one mask ≤ 10% of the image area).
Data were split in a stratified fashion: 90% of images for training, 10% for validation. Class imbalance overindividuals was mitigated with WaeightedRandomSampler.
Optimization employed SGD in three stages: (
After fine‑tuning, 2‑normalizedgallery and query embeddings were indexed in aFAISS IndexFlatIP [28, 29]; for each query the 50 nearest neighbors were retrieved, and confidence was defined as (cos +1)/2.
Descriptors from the three Dual‑backbone models complemented WildFusion and XGBoost inside the final meta‑algorithm, increasing ensemble robustness on challenging and rare cases and yielding an additional score gain (Ta2b).
a cascading scheme that invoked one to three models per species.
Eurasian lynx. (
Fire salamander. (
Loggerhead sea turtle. Dual‑backbone operates as the exclusive source: simil<ar7i0t.3y% is interpreted asnew_individual; otherwise the identifier proposed by the model is retained.
This species‑specific combination of WildFusion, XGBoost, and Dual‑backbone merges their errors that exhibit low mutual correlation. The ensemble achpireivveadteascore of 67.420% and a public score of 65.114%, securing second place among 172 teams in AnimalCLEF 2025 (T2a)b..
High‑precision Animal Re-ID methods open new opportunities in both scientific research and applied domains. These techniques facilitate automated census work and substantially facilitate field studies: instead of capturing and tagging animals, researchers can deploy camera traps, drones, or underwater cameras and then analyze the collected imagery algorithmically. Such approaches are already employed to monitor endangered species e.g., identifying snow leopards or whales allows estimating population size, migration routes, and individual longevity. Accurate and scalable Animal Re‑ID thus constitutes a key enabling factor of non‑invasive biodiversity monitoring.
Future improvements are envisioned along four complementary directions.
(i) Transductive graph‑based models. AnimalCLEF 2025 data exhibit a pronounced long‑tail distribution in the number of images pienrdividual. Under these circumstances, graph re‑ID strategies such as GCN-based reranking may redistribute confidence from majority classes to minority classes and raise recall in the tail of the distribu26t,io3n0].[ Initial GCN experiments reduced the public score; nevertheless, further exploration of neighborhood radii and regularization schemes remains promising.
(ii) Tiling and localized matching. Dividing images into distortion‑free squares and performing pairwise tile matching leads to a quadratic growth in computational complexity and GPU memory consumption, yet can mitigate background influence and raise confidence for fine‑scale spotted patterns.
(iii) Pseudo‑labeling and self‑training. Unlabeled images from GBIF (Global Biodiversity Information Facility) or surveillance video streams can augment the trainincognsfideetn.t Aset may be formed with current models, followed by additional backbone fine‑tuning while strictly controlling pseudo‑label accuracy.
(iv) Automatic threshold tuning. Bayesian optimization or diferentiable threshold tuning on the validation score would eliminate manual adjustment of the 39.7%, 13%, and 70.3% thresholds and adapt the meta‑algorithm to new species without manual threshold tuning [31].
The author is grateful to all individuals and organizations involved in the collection and annotation of data that enable the training of models and the development of animal re‑identification tools.
During the preparation of this work, the following Generative AI tool was employed: • ChatGPT o3 (OpenAI, June 2025 model) —Text Translation of the paper from Russian to English,
Grammar and spelling check, andImprove writing style.
All AI-generated suggestions were reviewed and edited manually; the authors assume full responsibility for the final content. No Generative AI system was used to create original scientific ideas, analyze data, or draw conclusions. [8] L. Picek, S. Kahl, H. Goëau, L. Adam, T. Larcher, C. Leblanc, M. Servajean, K. Janoušková, J. Matas, V. Čermák, K. Papafitsoros, R. Planqué, W.-P. Vellinga, H. Klinck, T. Denton, J. S. Cañas, G. Martellucci, F. Vinatier, P. Bonnet, A. Joly, Overview of lifeclef 2025: Challenges on species presence prediction and identification, and individual animal identification, in: International Conference of the Cross-Language Evaluation Forum for European Languages (CLEF), Springer, 2025. [9] L. Picek, E. Belotti, M. Bojda, L. Bufka, V. Cermak, M. Dula, R. Dvorak, L. Hrdy, M. Jirik, V. Kocourek, et al., Czechlynx: A dataset for individual identification and pose estimation of the eurasian lynx, arXiv preprint arXiv:2506.04931 (2025). [10] J. Deng, J. Guo, X. Niannan, S. Zafeiriou, Arcface: Additive angular margin loss for deep face recognition, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 4690–4699. do1i:0.1109/CVPR.2019.00482. [11] L. Adam, V. Cermak, K. Papafitsoros, L. Picek, Wildlifereid-10k: Wildlife re-identification dataset with 10k individual animals, in: Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, 2025, pp. 2099–2109. [12] Happywhale – whale and dolphin identification, Kaggle Competition, 2022. UhRttLp:s://www.
kaggle.com/competitions/happy-whale-and-dolphin. [13] Wild Me, LILA Science, Leopard id dataset, Dataset, 2022. UhtRtLp:s://lila.science/datasets/ leopard-id. [14] Wild Me, LILA Science, Hyena id dataset, Dataset, 2022. UhtRtLp: s://lila.science/datasets/hyen. a-id [15] L. Adam, V. Čermák, K. Papafitsoros, L. Picek, Seaturtleid2022: A long-span dataset for reliable sea turtle re-identification, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 7146–7156. [16] V. Čermák, L. Picek, L. Adam, K. Papafitsoros, Wildlifedatasets: An open-source toolkit for animal re-identification, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 5953–5963. [17] M. Tan, Q. V. Le, Eficientnetv2: Smaller models and faster training, in: M. Meila, T. Zhang (Eds.), Proceedings of the 38th International Conference on Machine Learning (ICML), volume 139 of Proceedings of Machine Learning Research, PMLR, 2021, pp. 10096–10106. URL: https://proceedings. mlr.press/v139/tan21a.html, long presentation. [18] J. Sun, Z. Shen, Y. Wang, H. Bao, X. Zhou, Loftr: Detector-free local feature matching with transformers, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8928–8937. URL: https://openaccess.thecvf.com/content/CVPR2021/html/Sun_LoFTR_ Detector-Free_Local_Feature_Matching_With_Transformers_CVPR_2021_paper.html. [19] P. Lindenberger, P.-E. Sarlin, M. Pollefeys, Lightglue: Local feature matching at light speed, arXiv preprint arXiv:2306.13643 (2023). URLh:ttps://arxiv.org/abs/2306.13643. [20] B. V. R. Alliance, Bvra/megadescriptor-l-3h8t4t, ps://huggingface.co/BVRA/MegaDescriptor-L-,384 2024. Model card on Hugging Face; accessed 2025-07-03. [21] D. DeTone, T. Malisiewicz, A. Rabinovich, Superpoint: Self-supervised interest point detection and description, 2018. URLh:ttps://arxiv.org/abs/1712.0762a9r.Xiv:1712.07629. [22] X. Zhao, X. Wu, W. Chen, P. C. Y. Chen, Q. Xu, Z. Li, Aliked: A lighter keypoint and descriptor extraction network via deformable transformation, 2023h.tUtRpLs://arxiv.org/abs/2304.03608. arXiv:2304.03608. [23] M. J. Tyszkiewicz, P. Fua, E. Trulls, Disk: Learning local features with policy gradient, 2020. URL: https://arxiv.org/abs/2006.1356a6r.Xiv:2006.13566. [24] D. G. Lowe, Distinctive image features from scale–invariant keypoints, International Journal of
Computer Vision 60 (2004) 91–110. do1i:0.1023/B:VISI.0000029664.99615.94. [25] WildlifeDatasets, wildlife-tools: Reference implementation of wildhftutsiposn:/,/github.com/
WildlifeDatasets/wildlife-tools, 2024. Accessed 2025-07-03. [26] Z. Zhong, L. Zheng, D. Cao, S. Li, Re-ranking person re-identification with k-reciprocal encoding, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4656–4665. URL: https://openaccess.thecvf.com/content_cvpr_2017/html/Zhong_Re-Ranking_ Person_Re-Identification_CVPR_2017_paper.html. [27] H. Wang, Z. Du, M. Du, F. Yang, S. Hu, S. Liu, J. Zhou, X. Hu, Score‑cam: Score‑weighted visual explanations for convolutional neural networks, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPR W), IEEE, 2020, pp. 111–119. doi:10.1109/CVPRW50498.2020.00020. [28] M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.-E. Mazaré, M. Lomeli, L. Hosseini,
H. Jégou, The faiss library (2024a).rXiv:2401.08281. [29] J. Johnson, M. Douze, H. Jégou, Billion-scale similarity search with GPUs, IEEE Transactions on
Big Data 7 (2019) 535–547. [30] Y. Zhang, Q. Qian, H. Wang, C. Liu, W. Chen, F. Wang, Graph convolution based eficient re-ranking for visual retrieval, 2023. URhLt:tps://arxiv.org/abs/2306.0879a2r.Xiv:2306.08792. [31] J. Snoek, H. Larochelle, R. P. Adams, Practical bayesian optimization of machine learning algorithms, in: Advances in Neural Information Processing Systems (NeurIPS), 2012, pp. 2951–2959. URL: https://proceedings.neurips.cc/paper_files/paper/2012/file/ 05311655a15b75fab86956663e1819cd-Paper.pdf.