The first edition of the individual animal identification challenge, AnimalCLEF 2025, organized within LifeCLEF, advances the field of animal re-identification using computer vision and machine learning. Building on the WildlifeReID-10k dataset and incorporating new data, AnimalCLEF 2025 challenges participants to recognize individual animals from images for three species: lynxes, salamanders and sea turtles. The mix of species with diferent image capture conditions attempts to make the submitted prediction models generalizable to unseen species. The competition attracted 270 participants across 230 teams, with 136 outperforming the provided baseline based on MegaDescriptor. This overview paper provides (i) a comprehensive description of the challenge and provided baseline method, (ii) detailed characteristics of the dataset and task specifications, (iii) an examination of the methods employed by contestants, and (iv) a discussion of the competition outcomes. The results highlight incremental advancements in animal re-identification, showcasing innovative approaches and techniques that push the limits of previous work.
Animal re-identification is crucial for studying various aspects of wildlife, including disease monitoring
and control [
To address these issues and enable continual incremental improvements in animal re-identification, we
organize the AnimalCLEF research competition. This first edition, which took part in the LifeCLEF 2025
[
By promoting innovation in this field, AnimalCLEF 2025 aims to bridge further the gap between computer vision capabilities and practical conservationist needs, potentially impacting areas ranging from biodiversity research to habitat protection and beyond.
The AnimalCLEF 2025 dataset builds upon the previous editions of LifeCLEF [15–17] and the
WildlifeReID-10k dataset [
In addition to the training set, we assembled a test dataset with images never seen and shared on the internet, including three diferent species, i.e., Lynx lynx, Salamandra salamandra, and Caretta caretta. The lynx data is part of the newly released dataset, CzechLynx [48]. The turtle data is from an unreleased part of the SeaTurtleID2022 dataset [19]. The salamander data has not yet been released. All three datasets were split into the database and query sets, which are briefly described below, and their statistics in detail are listed in Table 1.
The subset with Eurasian lynxes (Lynx lynx) individuals includes real camera-trap photographs and synthetic samples, organized around three computer-vision tasks: individual identification, pose estimation, and instance segmentation. It spans more than 15 years and covers two regions in Central Europe (Southwest Bohemia and the Western Carpathians) and contains around 37k manually verified and labeled images. The images used for the AnimalCLEF 2025 challenge all contain real animals. The splits were provided by the authors [48].
The subset with fire salamander (Salamandra salamandra) individuals consists of images collected under specific environmental conditions. As these amphibians are primarily active at night and only under high humidity, most captures occurred during rainy nights. Within a fixed area, salamanders were located using visual encounter surveys conducted by a single, consistent observer throughout the study. Each individual was then photographed from dorsal, ventral, and right lateral perspectives. Following standard protocol, all salamanders were released immediately at the site of capture without being marked.
The subset with sea turtles (Caratta carreta) individuals contains 8,729 photographs of 438 individuals collected within 13 years at Zakynthos Island, Greece. Each photograph includes various annotations, e.g., identity, encounter timestamp, and body parts segmentation masks. As the database, we considered the 8,729 images from the original publication [19]. As the query set, we additionally published 500 images of 100 individuals, 50 of which were already present in the database and 50 were new.
For all query images, the competitors needed to decide whether the depicted individual is known (present in the database). If yes, they additionally needed to specify the identity (label) of the individual. Therefore, the task was essentially open-set. The training set was completely optional, allowing trading fast training for accuracy.
The evaluation was performed by accuracy on known and unknown samples. The accuracy was further balanced with respect to datasets and individuals, preventing overrepresented individuals from having too big an impact on the solution. Mathematically:
BAKS = 1 ∑︁ =1 |1| ∑∈︁ |1,| ∈∑︁, 1(ˆ = ),
BAUS = 1 ∑︁ 1
∑︁ =1 |* | ∈* |
1 ,| ∈∑︁, 1(ˆ = new), (1) (2) where 1 is the indicator (0/1) function, = 3 the number of datasets, (or * ) the known (or unknown) individuals (classes) in dataset and , indices of all images of individual in the query set of the dataset . Finally, and ˆ are the true and predicted identities of the -th image.
To evaluate overall performance with a single score, we combine the closed-set (BAKS) and open-set (BAUS) metrics using their geometric mean:
normalized accuracy = √BAKS · BAUS.
The geometric mean is preferred over the arithmetic mean, as it penalises extreme imbalances, e.g., a trivial classifier predicting only new individuals yields BAKS = 0% and BAUS = 100%, resulting in a misleading 50% arithmetic mean but a more appropriate geometric mean of 0%.
To enable an easier start for all participants and straightforward model evaluation, we provide two
baselines based on the pre-trained models. The first one was the MegaDescriptor-L-384 [
Both models achieved relatively strong performance in animal re-identification, serving as a robust starting point. Overall, our goal was to ofer a comprehensive and accessible starting point for all participants, enabling them to focus on developing novel solutions and improving upon the baselines.
The AnimalCLEF 2025 competition was launched on February 28, 2025, and was promoted through the LifeCLEF [50], Kaggle, and FGVC challenge web pages, inviting participants to register. The competition ran for approximately three months, with the final submission deadline on May 19. Similar to other LifeCLEF challenges, the test data remained confidential. Participants were allowed to make up to ifve submissions per day using the Kaggle evaluation platform to assess their models on the public query set. Based on these results, the participants were allowed to select two submissions, which were evaluated on the private query set. The participants were scored and ranked based on the better of the two selected solutions.
Participants were strongly encouraged to submit both their code and a detailed technical report (Working Notes) to ensure their results can be fully reproduced. All the submitted Working Notes underwent thorough review and were given complex feedback by 2-3 experts with extensive publication records in Computer Vision and Machine Learning. This rigorous review process was designed to guarantee reproducibility and maintain quality standards. The review was single-blind, allowing participants to respond with up to two rebuttals to address any feedback or concerns raised by the reviewers. These working notes provide an in-depth analysis of the techniques employed, including hyperparameter tuning, model ensembling, and loss function selection, ofering valuable insights into the development of the method for animal re-identification.
270 participants in 230 teams participated, with 136 teams outperforming the MegaDescriptor baseline (with normalized accuracy 0.309) and 88 teams outperforming the MegaDescriptor+ALIKED baseline (with normalized accuracy 0.443). The provided baseline was very good at detecting new individuals (BAUS of 0.992), while it had problems correctly classifying the known individuals (BAKS of 0.209).
From the Kaggle platform, we were able to extract all submissions but not the two final submissions that the participants selected for the final evaluation as described in Section 2.4. To analyze the participants’ performance for this paper, we selected their submission with the highest public score and evaluated it on the private leaderboard. Therefore, the analysis presented here may slightly difer from the oficial leaderboard. The ranking of teams is taken from the oficial ranking.
Figure 2 (top) shows the results of the best 50 participants. All results are shown on the private dataset. The dash-dotted lines depict the normalized accuracy on individual datasets, while the full blue is the overall normalized accuracy computed as their mean. The normalized accuracy on sea turtles (dash-dotted orange) is relatively stable for all teams. The ranking of teams was most influenced by their performance on lynxes (dash-dotted red) and salamanders (dash-dotted black). Interestingly, the second-ranked team (webmaking) achieved the best performance on salamanders and sea turtles but was ranked 25 on lynxes. The winning team (DataBoom) achieved a more consistent performance on all datasets. If the results of these two best-performing teams were combined, the accuracy would increase from 0.713 (DataBoom) or 0.675 (webmaking) to 0.734.
Figure 2 (bottom) shows BAKS and BAUS averaged over all datasets. These numbers show that all teams achieved very high BAUS, that is, they predicted new individuals very well. We also show precision on unknown samples, defined as the ratio of correctly predicted unknowns over all predicted unknowns. Since the precision is between 0.5 and 0.6 and BAKS between 0.9 and 1.0, that means that all participants predicted approximately twice the number of unknown individuals than the true number of unknown individuals and that these predictions contained almost all unknown individuals.
We have received eight submissions, all of which outperformed the baseline MegaDescriptor, and six outperformed the WildFusion baseline. Six submissions were accepted into the proceedings. Details of the best methods and systems used are synthesised below and further developed in participants’ working notes [51–56].
Team DataBoom [51] (Top 1) improved on WidlFusion by using multiple local feature-matching techniques and by replacing MegaDescriptor with MiewID. They used MiewID as a preselected which did not enter the final score computation. The scores from local feature-matching techniques are combined by averaging. The authors also argued that some photos must be properly segmented to remove artifacts such as human fingers for salamanders.
Team webmaking [52] (Top 2) used two streams: (i) WildFusion, which provides calibrated global–local similarity scores, followed by two reranking methods applied to the top-20 images, and (ii) an XGBoost model trained on concatenated features from MegaDescriptor and MIEW-ID. The authors tuned the weights of both streams for all three datasets and provided additional experiments with preprocessing contour masks, image alignment and background blurring, which, however, did not bring a significant gain in accuracy.
Team Tim Riggins [53] (Top 7) combined the WildFusion baseline with the gradient boosting method LightGBM [57], which is an inexact fast approximation of the conventional Gradient Boosting Decision Tree (GBDT) method [58]. The authors did not perform any fine-tuning of the feature extractor, making their algorithm eficient.
Team already mygo [54] (Top 34) used a similar approach to the WildFusion baseline, thus a combination of deep method (Swin) with local features (DISK and ALIKED) combined via the WildFusion method. The authors used fine-tuning of the feature extractor on the database and performed an analysis of the thresholding parameter used to classify new individuals. Team dongyeonkim10 [55] (Top 44) used data preprocessing techniques and then averaged the WildFusion score with EVA02 [59] score. The authors also investigated inference-time augmentations and included techniques which did not work for them.
Team DS@GT LifeCLEF [56] (Top 103) compares features extracted by DINOv2 and MegaDescriptor-L, reprojected using triplet loss, and classified using a k-nearest neighbors (kNN) classifier. Since they ifne-tuned the baseline MegaDescriptor, it outperforms it, but falls short of top leaderboard entries. 0.8 cy0.7 a r u cca0.6 d e z lia0.5 m r o n0.4
1 0.8 y c rau0.6 c c a 0.4 0.2 lynxes salamanders sea turtles
combined known: BAKS unknown: BAUS unknown: precision ooBmiagkn zzhhh lianAmsaanH itsaunV iisggnR a666gyn agguhnY izzaxgyn euhZX lllkLhhh trseeon taabddy ISLAD rscekn lkuHiitcgBD liieaboodRNmfilirsseannKC izggooLuuhn iiicbboohnCmeauhnwllceebodkyuuh teeayuhPnnBT iJrraodnCMM reuUww IEANWirraonm rrsoAw liceANlieanD tsgooo goymlepuV ltaa1n liitsgghn lliseeaonnD lffteaSonWagohhnnTC faaSoyyLnn littszeaaa_vvoE tJrceeeagdkhE ie01goodkynnm llenAlroyG aagdA ljakph 193 ebv 1D lou taaDebwm Ion iTmiijagqnn ianX am MeoddunB en Wljeaadn liseyE le lreaady fteaSn Gh N Su leaomcehO raanK A S_BUGM n A
Z an j e l
A team name (ranked) v e P _ 5 2 3 _ U S M d h A p e s o J
This paper presents an overview and evaluation of the results of the AnimalCLEF 2025 challenge, organized in conjunction with the LifeCLEF and CVPR-FGVC12 workshops. The challenge introduced a multi-species, open-set re-identification task using the WildlifeReID-10k dataset for training and new identities from three species, e.g., Caretta caretta, Salamandra salamandra, and Euroasian lynx, for testing. Participants were tasked with developing a model that automatically determines whether a query image belongs to a known individual in the gallery set or represents a new, unseen individual. The task emphasized robustness under real-world conditions, including background clutter, low sample counts, and interspecific variation. The competition attracted over 230 teams and demonstrated a wide variety of methodological approaches. Notably, global feature extractors, such as MegaDescriptor, remained central, while a resurgence in local matching and calibrated score fusion strategies highlighted the importance of fine-grained cues. Here are the key takeaways from this edition: • Global-local fusion is the new baseline. Nearly all top-performing teams employed hybrid methods combining global embeddings (e.g., MegaDescriptor, MIEW) with local feature matchers (e.g., ALIKED, DISK, SuperPoint). These were fused using either WildFusion or custom weighted schemes, leading to significant gains over global-only systems. • Score calibration is essential in open-set recognition. Teams that succeeded in balancing BAKS and BAUS scores commonly applied threshold-based novelty detection combined with isotonic or learned calibration of similarity scores. Without this, methods either overpredicted new individuals or failed to detect them altogether. • Diversity in global embedding sources boosts generalization. Teams leveraging multiple descriptor models (e.g., Swin-L from MegaDescriptor and CNN-based MIEW or EVA02) demonstrated that architectural diversity leads to more robust predictions across species and conditions. • Meta-learning and ensemble stacking improve robustness. The runner-up solution implemented a meta-algorithm with staged fusion: global-local WildFusion scores fed into an XGBoost classifier, followed by ArcFace fine-tuning per species. This approach showed cumulative improvements at each stage.
Directions for Future Editions. The first edition of AnimalCLEF established a strong foundation for evaluating open-set animal re-identification. Looking ahead, we identify several key directions that could enrich future editions and better reflect real-world biodiversity monitoring needs: 1. Temporal robustness and longitudinal tracking. Extending the task to multi-year observations (e.g., from juvenile to adult) would test models’ ability to handle appearance changes over time; a critical requirement in wildlife research. 2. Species-agnostic or cross-species modeling. Unified models that generalize across multiple or unseen species would reflect real-world deployments, especially in diverse environments. 3. Identity and category discovery. Introducing unsupervised or semi-supervised settings where new individuals or even species must be grouped or discovered would push the boundaries of open-world recognition. 4. Multimodal and metadata-aware tasks. Although image-only systems dominated in 2025, future tracks could explicitly encourage the use of spatial, temporal, and ecological metadata to enrich recognition and enable more contextual reasoning.
We believe these extensions would not only stimulate innovation in machine learning but also bring AnimalCLEF closer to practical conservation and ecological monitoring scenarios.
During the preparation of this work, the authors used Grammarly for grammar and spelling checks and ChatGPT for improving clarity and rewording sentences. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.
LA has been supported by the Ministry of Education, Youth and Sports of the Czech Republic under the project SGS-2024-017. with calibrated similarity fusion, in: European Conference on Computer Vision, Springer, 2025, pp. 18–36. [15] T. Lorieul, E. Cole, B. Deneu, M. Servajean, A. Durso11, I. Bolon, H. Glotin, R. Planqué, R. R. de Castaneda, W.-P. Vellinga, et al., Overview of lifeclef 2021: An evaluation of machine-learning based species identification and species distribution prediction, in: Experimental IR Meets Multilinguality, Multimodality, and Interaction: 12th International Conference of the CLEF Association, CLEF 2021, Virtual Event, September 21–24, 2021, Proceedings, volume 12880, Springer Nature, 2021, p. 371. [16] A. Joly, H. Goëau, S. Kahl, B. Deneu, M. Servajean, E. Cole, L. Picek, R. R. De Castaneda, I. Bolon, A. Durso, et al., Overview of lifeclef 2020: a system-oriented evaluation of automated species identification and species distribution prediction, in: International Conference of the CrossLanguage Evaluation Forum for European Languages, Springer, 2020, pp. 342–363. [17] A. Joly, H. Goëau, S. Kahl, L. Picek, C. Botella, D. Marcos, M. Šulc, M. Hrúz, T. Lorieul, S. S.
Moussi, et al., Lifeclef 2023 teaser: Species identification and prediction challenges, in: European Conference on Information Retrieval, Springer, 2023, pp. 568–576. [18] L. Adam, K. Papafitsoros, C. Jean, A. Rees, Exploiting facial side similarities to improve ai-driven sea turtle photo-identification systems, bioRxiv (2024) 2024–09. [19] 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 (WACV), 2024, pp. 7146–7156. [20] W. Andrew, J. Gao, S. Mullan, N. Campbell, A. W. Dowsey, T. Burghardt, Visual identification of individual holstein-friesian cattle via deep metric learning, Computers and Electronics in Agriculture 185 (2021) 106133. URL: https://www.sciencedirect.com/science/article/pii/S0168169921001514. [21] W. Andrew, C. Greatwood, T. Burghardt, Visual localisation and individual identification of holstein friesian cattle via deep learning, in: Proceedings of the IEEE International Conference on Computer Vision Workshops, 2017, pp. 2850–2859. URL: https://openaccess.thecvf.com/content_ ICCV_2017_workshops/w41/html/Andrew_Visual_Localisation_and_ICCV_2017_paper.html. [22] W. Andrew, S. Hannuna, N. Campbell, T. Burghardt, Automatic individual holstein friesian cattle identification via selective local coat pattern matching in RGB-D imagery, in: 2016 IEEE International Conference on Image Processing (ICIP), IEEE, 2016, pp. 484–488. URL: https:// ieeexplore.ieee.org/abstract/document/7532404. [23] J. Bruslund Haurum, A. Karpova, M. Pedersen, S. Hein Bengtson, T. B. Moeslund, Re-identification of zebrafish using metric learning, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, 2020, pp. 1–11. URL: https://ieeexplore.ieee.org/ document/9096922. [24] Local Ocean Conservation, Turtle recall: Conservation challenge, 2021. URL: https://zindi.africa/ competitions/turtle-recall-conservation-challenge. [25] N. Dlamini, T. L. van Zyl, Automated identification of individuals in wildlife population using siamese neural networks, in: 2020 7th International Conference on Soft Computing & Machine Intelligence (ISCMI), IEEE, 2020, pp. 224–228. [26] A. C. Ferreira, L. R. Silva, F. Renna, H. B. Brandl, J. P. Renoult, D. R. Farine, R. Covas, C. Doutrelant, Deep learning-based methods for individual recognition in small birds, Methods in Ecology and Evolution 11 (2020) 1072–1085. [27] A. Freytag, E. Rodner, M. Simon, A. Loos, H. S. Kühl, J. Denzler, Chimpanzee faces in the wild: Log-Euclidean CNNs for predicting identities and attributes of primates, in: German Conference on Pattern Recognition, Springer, 2016, pp. 51–63. [28] L. Fu, G. He, Cow dataset, 2021. URL: https://doi.org/10.6084/m9.figshare.16879780.v1. [29] J. Gao, T. Burghardt, W. Andrew, A. W. Dowsey, N. W. Campbell, Towards self-supervision for video identification of individual holstein-friesian cattle: The Cows2021 dataset, arXiv preprint arXiv:2105.01938 (2021). URL: https://arxiv.org/abs/2105.01938. [30] Z. He, J. Qian, D. Yan, C. Wang, Y. Xin, Animal re-identification algorithm for posture diversity, in: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2023, pp. 1–5. [31] J. Holmberg, B. Norman, Z. Arzoumanian, Estimating population size, structure, and residency time for whale sharks Rhincodon typus through collaborative photo-identification, Endangered Species Research 7 (2009) 39–53. URL: https://www.int-res.com/abstracts/esr/v7/n1/p39-53/. [32] D. Kern, T. Schiele, U. Klauck, W. Ingabire, Towards automated chicken monitoring: Dataset and machine learning methods for visual, noninvasive reidentification, Animals 15 (2024) 1. [33] M. Lahiri, C. Tantipathananandh, R. Warungu, D. I. Rubenstein, T. Y. Berger-Wolf, Biometric animal databases from field photographs: identification of individual zebra in the wild, in: Proceedings of the 1st ACM international conference on multimedia retrieval, 2011, pp. 1–8. [34] S. Li, J. Li, H. Tang, R. Qian, W. Lin, ATRW: A Benchmark for Amur Tiger Re-identification in the Wild, in: Proceedings of the 28th ACM International Conference on Multimedia, Association for Computing Machinery, 2020, p. 2590–2598. [35] T.-Y. Lin, Cat individual images, 2020. URL: https://www.kaggle.com/datasets/timost1234/ cat-individuals. [36] V. Miele, G. Dussert, B. Spataro, S. Chamaillé-Jammes, D. Allainé, C. Bonenfant, Revisiting animal photo-identification using deep metric learning and network analysis, Methods in Ecology and Evolution 12 (2021) 863–873. [37] G. Mougeot, D. Li, S. Jia, A deep learning approach for dog face verification and recognition, in: PRICAI 2019: Trends in Artificial Intelligence: 16th Pacific Rim International Conference on Artificial Intelligence, Cuvu, Yanuca Island, Fiji, August 26-30, 2019, Proceedings, Part III 16, Springer, 2019, pp. 418–430. [38] E. Nepovinnykh, T. Eerola, V. Biard, P. Mutka, M. Niemi, M. Kunnasranta, H. Kälviäinen, Sealid:
Saimaa ringed seal re-identification dataset, Sensors 22 (2022) 7602. [39] J. R. Parham, J. Crall, C. Stewart, T. Berger-Wolf, D. Rubenstein, Animal population censusing at scale with citizen science and photographic identification, in: 2017 AAAI Spring Symposium Series, 2017. URL: https://www.aaai.org/ocs/index.php/SSS/SSS17/paper/viewPaper/15245. [40] C. Trotter, G. Atkinson, M. Sharpe, K. Richardson, A. S. McGough, N. Wright, B. Burville, P. Berggren, NDD20: A large-scale few-shot dolphin dataset for coarse and fine-grained categorisation, arXiv preprint arXiv:2005.13359 (2020). URL: https://arxiv.org/abs/2005.13359. [41] O. Wahltinez, S. J. Wahltinez, An open-source general purpose machine learning framework for individual animal re-identification using few-shot learning, Methods in Ecology and Evolution 15 (2024) 373–387. [42] L. Wang, R. Ding, Y. Zhai, Q. Zhang, W. Tang, N. Zheng, G. Hua, Giant panda identification, IEEE Transactions on Image Processing 30 (2021) 2837–2849. URL: https://ieeexplore.ieee.org/document/ 9347819. [43] WildMe, Beluga ID 2022, 2022. URL: https://lila.science/datasets/beluga-id-2022. [44] WildMe, Leopard ID 2022, 2022. URL: https://africancarnivore.wildbook.org/. [45] P. Yu, T. Burghardt, A. W. Dowsey, N. W. Campbell, Multicamcows2024–a multi-view image dataset for ai-driven holstein-friesian cattle re-identification on a working farm, arXiv preprint arXiv:2410.12695 (2024). [46] M. Zuerl, R. Dirauf, F. Koeferl, N. Steinlein, J. Sueskind, D. Zanca, I. Brehm, L. v. Fersen, B. Eskofier, Polarbearvidid: A video-based re-identification benchmark dataset for polar bears, Animals 13 (2023) 801. [47] S. Zufi, A. Kanazawa, T. Berger-Wolf, M. J. Black, Three-D safari: Learning to estimate zebra pose, shape, and texture from images "In the wild", in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 5359–5368. URL: https://ieeexplore.ieee.org/document/ 9010937. [48] L. Picek, E. Belotti, M. Bojda, L. Bufka, V. Cermak, M. Dula, R. Dvorak, L. Hrdy, M. Jirik, V. Kocourek, J. Krausova, J. Labuda, J. Straka, L. Toman, V. Trulık, M. Vana, M. Kutal, Czechlynx: A dataset for individual identification and pose estimation of the eurasian lynx, arXiv preprint arXiv:1804.07177 (2025). [49] Z. Liu, H. Hu, Y. Lin, Z. Yao, Z. Xie, Y. Wei, J. Ning, Y. Cao, Z. Zhang, L. Dong, et al., Swin transformer v2: Scaling up capacity and resolution, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 12009–12019. [50] A. Joly, L. Picek, S. Kahl, H. Goëau, L. Adam, C. Botella, M. Servajean, D. Marcos, C. Leblanc, T. Larcher, et al., Lifeclef 2025 teaser: Challenges on species presence prediction and identification, and individual animal identification, in: European Conference on Information Retrieval, Springer, 2025, pp. 373–381. [51] S. Lanskikh, G. Demidov, D. Dinmuhametov, A. Khlopotnukh, K. Bogdan, R. Pakhomov, Matchers with weights, in: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, 2025. [52] N. Semenova, Dual-stream meta-algorithm for animal re-id: Wildfusion global-local matching and xgboost on megadescriptor-miew embeddings, in: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, 2025. [53] S. Fedorchenko, S. Arefiev, Gradient boosting similarity in entity matching, in: Working Notes of
CLEF 2025 - Conference and Labs of the Evaluation Forum, 2025. [54] J. Tan, A. Wang, Open-set animal re-identification via multilevel feature fusion, in: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, 2025. [55] D. Kim, B. Park, H. Bae1, S. Lee, C. Lee, Fusion of global and local descriptors with feature calibration for multi-species animal re-identification: Animalclef 2025, in: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, 2025. [56] A. Miyaguchi, C. Maruthaiyannan, C. R. Clark, Ds@gt animalclef: Triplet learning over vit manifolds with nearest neighbor classification for animal re-identification, in: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum, 2025. [57] G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, T.-Y. Liu, Lightgbm: A highly eficient gradient boosting decision tree, Advances in neural information processing systems 30 (2017). [58] J. H. Friedman, Greedy function approximation: a gradient boosting machine, Annals of statistics (2001) 1189–1232. [59] Y. Fang, Q. Sun, X. Wang, T. Huang, X. Wang, Y. Cao, Eva-02: A visual representation for neon genesis, Image and Vision Computing 149 (2024) 105171.