Animal re-identification allows for non-invasive, scalable monitoring of wildlife by matching visual cues specific to each individual. In this paper, we present a mapping-oriented, open system that progressively filters and validates candidate images. We took the basic step-by-step solution presented in the competition and improved each step in it. First, we calculate global attachments to select a small subset of potentially suitable images for each query. Then we use a set of local feature matching tools, each of which contains a separate detector-descriptor pair and a matching algorithm, to obtain additional similarity estimates that capture the smallest visual details, such as unique markings and morphological patterns. Then we combine these estimates for each user using a studied, weighted synthesis mechanism that identifies the most reliable features for diferent types and shooting conditions. Finally, a calibrated confidence threshold allows you to separate previously seen individuals from new ones, ensuring reliable recognition when new animals are detected. We are evaluating based on the AnimalCLEF 2025 collections (loggerhead sea turtles, salamanders, and Eurasian lynx), and our system provides highly balanced accuracy for both known and unknown classes. The modular design makes it easy to adapt additional mapping devices or embedded models, demonstrating resistance to background interference, blockages, and variable shooting conditions. With this solution, we took the first place in the AnimalCLEF2025 competition.
Animal re-identification is a critical task in wildlife research and conservation, enabling the tracking of individuals over time to study population dynamics, habitat use, migration patterns, and behavior. By recognizing unique traits—such as markings, color patterns, or morphological features—researchers can monitor species in a non-invasive, scalable manner. Automating this process not only accelerates data collection but also enhances the consistency and scale at which individuals can be reliably tracked. These capabilities are vital for identifying biodiversity threats and supporting evidence-based conservation strategies.
Despite recent progress in computer vision and machine learning, reliably identifying individual animals remains challenging. Models often overfit to environmental cues such as background, lighting, or camera angle—rather than focusing on species-specific, individual characteristics. This results in poor generalization to new environments or image conditions, limiting the practical efectiveness of many re-identification systems in real-world conservation settings.
The AnimalCLEF 2025 challenge, part of the LifeCLEF 2025 evaluation campaign, addresses this
problem through the task of individual animal identification for three wildlife species: loggerhead sea
turtles (Caretta caretta [
© 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
in the reference database) or is a new, previously unseen individual. If known, the correct identity
must be assigned. To aid generalization, participants are allowed to augment their models using the
WildlifeReID-10k dataset—a large-scale benchmark comprising over 140,000 images across 10,000+
individuals from a diverse set of species [
In this technical report, we present our solution to the AnimalCLEF 2025 challenge. Our pipeline consists of four core stages: global candidate selection, local visual matching, score aggregation (bagging), and novelty filtering. This modular structure combines coarse-to-fine similarity evaluation with a confidence-based thresholding mechanism to distinguish between known and novel individuals. Our approach balances precision and generalization while remaining robust across the three target species.
Several prior works address fine-grained wildlife classification and open-set identification. CNN
ensembles and metadata fusion prove efective for discriminating visually similar species. Local feature
matchers (e.g., SuperPoint[
To properly evaluate the performance of our re-identification pipeline on both known and novel individuals, we employ three metrics: Balanced Accuracy on Known Samples (BAKS), Balanced Accuracy on Unknown Samples (BAUS), and their Geometric Mean (GeoMean).
The Balanced Accuracy on Known Samples (BAKS) quantifies the model’s performance on individuals that are present in the training dataset. Unlike standard accuracy, BAKS is computed in a class-balanced manner to mitigate the efects of class imbalance.
Let be the set of known classes. Then:
BAKS = 1 ∑︁
TP || ∈ TP + FN , where: where: nosep TP is the number of true positive predictions for unknown class , nosep FN is the number of false negatives for class . nosep TP is the number of true positive predictions for class , nosep FN is the number of false negatives for class .
The Balanced Accuracy on Unknown Samples (BAUS) measures the model’s ability to correctly recognize individuals from unseen classes, i.e., classes not present in the training dataset.
Let be the set of unknown classes. Then:
BAUS = 1 ∑︁
TP | | ∈ TP + FN
To combine performance on both known and unknown samples, we compute the geometric mean of BAKS and BAUS:
GeoMean = √BAKS × BAUS.
Efective individual identification in wildlife datasets is often hindered by visual noise, background clutter, and inconsistencies in object localization. These issues are particularly prominent in the AnimalCLEF2025 dataset, where species such as salamanders are often partially occluded by human hands, and segmentation annotations are absent for several species (notably salamanders and sea turtles). To address this, we implement a targeted data preprocessing strategy centered on segmentation-aware data augmentation. For sea turtles, for instance, the ocean background in most images naturally isolates the subject, efectively acting as an implicit segmentation mask. Lynxes are already segmented.
We construct a supplementary annotated dataset[
The fine-tuned YOLOv11m-seg model significantly improves the localization quality of salamander regions by learning to suppress irrelevant background areas, particularly human hands and other visual artifacts. Qualitative inspection confirms more compact and accurate segmentation masks, which are then used to crop or mask the input images during downstream embedding extraction.
The implemented segmentation strategy significantly reduces the negative impact of background noise in species where it is dominant. Most notably, the use of segmentation for salamanders yields a measurable improvement in downstream matching metrics, confirming the hypothesis that accurate localization of the individual is a crucial prerequisite for reliable embedding computation and similarity-based identification.
Our baseline system integrates global deep descriptors and local keypoint-based matching to address open-set individual identification. The implementation relies on the wildlife-datasets and wildlife-tools libraries, designed to facilitate data loading, feature extraction, and similarity computation for wildlife-related tasks. Throughout this work, the term baseline refers to the publicly available solution Baseline with WildFusion.
The pipeline consists of the following components: Our pipeline extracts global descriptors using the
MegaDescriptor-L-384[
Global features are extracted from a high-capacity vision transformer model MegaDescriptor-L-384, pre-trained and fine-tuned to generate robust embeddings. These embeddings capture coarse-grained appearance information and are compared using cosine similarity.
Given two embeddings e, e ∈ R, cosine similarity is defined as: simcos(e, e) =
e · e . ‖e‖ ‖e‖
The baseline uses ALIKED as the local keypoint detector and descriptor. It is designed for accurate and eficient keypoint extraction with descriptor computation. The extractor outputs dense descriptors d and confidence scores.
These local descriptors are matched using MatchLightGlue, an attention-based matcher inspired by LightGlue. Let denote the attention afinity between descriptor in image and in image : After filtering matches via mutual nearest neighbors, the local similarity score ℓ(, ) is: exp(Wd · Wd ) = ∑︀′ exp(Wd · Wd′
) . ℓ(, ) = 1
∑︁ |ℳ(, )| (,)∈ℳ(,) .
The WildFusion[
Fused similarity scores are passed through isotonic regression, fitting a monotonic mapping to minimize squared error between calibrated scores and ground truth labels.
The basic level has a number of limitations. Initially, the approach presents only one matching method, the embedding model may not suit all species equally, the similarity matrices are averaged, and the lack of initial segmentation of salamanders creates background noise.
The proposed solution encompasses a four-stage process: Identify top- global embedding matches.
Compare candidates using multiple extractor–matcher pairs. Combine per-pair scores via learned weights ,. Apply threshold to calibrated scores to flag new individuals.
For each test sample , we compute its embedding e ∈ R using the pre-trained embedding model. Denote the set of all training embeddings as {etrain}=1. We compute cosine distances: We then select the top- training samples with the smallest distances: cos(︀ e, etrain)︀ = 1 −
e · etrain ‖e‖ ‖etrain‖ . () = arg min top{︀ cos(e, etrain)}︀ =1.
These candidates form the candidate set for local matching.
We have replaced MegaDescriptor-L-384 with a higher-quality miewid-msv3 [
Each test image is compared against each candidate ∈ () using several “extractor + matcher” pairs. Let
ℰ = {KeyNetAfNetHardNet , DISK (Kornia[
ℳ = {AdaLAM, GADMatcher, MatchLightGlue} be the set of matchers. For each extractor ∈ ℰ and matcher ∈ ℳ, we compute a local similarity matrix:
,(, ) = LocalSim,(, ), where LocalSim, denotes the local matching procedure with extractor and matcher . Each entry of the local similarity matrix corresponds to a confidence-weighted sum of matched keypoints.
The selection of the Kornia library was motivated by its provision of models capable of handling afine transformations, which expanded the comparison capabilities beyond the single matcher (MatchLightGlue) and limited set of extractors (SIFT, SuperPoint, ALIKED, DISK) available in wildlife_tools.
We replaced the ineficient averaging of the proximity matrix with a weighted vote.
Let ,(, ) be the similarity score for a given pair (, ) obtained from the local similarity matrix corresponding to each (, ).
We introduce a set of coeficients, ,, where each coeficient is associated with a unique combination of and . These coeficients allow us to assign diferent weights to the similarity scores from each local matrix. The final similarity score (, ) is then computed as a weighted mean of these per-pair scores: (, ) =
The optimal values for these coeficients, ,, are determined through an optimization process. This optimization is performed on a training dataset using scipy.optimize.minimize[14] with the Powell method. The objective of this optimization is to minimize the geomin metric as defined by Baks and Baus, ensuring that the final similarity score (, ) is well-calibrated for the given task.
Based on the aggregated similarity agg(, ), the most probable entity (predicted individual) and its corresponding confidence score are identified for each test image. To ascertain new individuals (i.e., those not present in the training dataset), a confidence threshold( ) is established. The threshold is chosen to maximize GeoMean on the validation set: and assign:
⎧⎨arg max (, ), if max() ≥ ,
∈ ()
⎩new_individual, otherwise.
* = arg max √︀BAKS( ) ×
∈ [
Score
We explore alternative embeddings (CLIP[16], DinoV2[17]) and dense matchers (LoFTR[18]). While these show promise, our species-specific pipeline remains most consistent. Future work includes species-routing hybrid matchers for further gains.
We thank the LifeCLEF organizers for the AnimalCLEF 2025 datasets and the WildlifeReID-10k auxiliary data. We also acknowledge developers of wildlife-datasets and wildlife-tools.
In preparing this work, the authors used GPT-4o to check grammar and spelling. In addition, the authors use GPT-4o to translate text into English. After using these tools/services, the authors reviewed and edited the content as needed and are solely responsible for the content of the publication. [14] P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey, İ. Polat, Y. Feng, E. W. Moore, J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Henriksen, E. A. Quintero, C. R. Harris, A. M. Archibald, A. H. Ribeiro, F. Pedregosa, P. van Mulbregt, SciPy 1.0 Contributors, SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python, Nature Methods 17 (2020) 261–272. doi: 10.1038/ s41592-019-0686-2. [15] The solution is available online at https://github.com/XXXM1R0XXX/AnimalCLEF2025, ???? [16] G. Ilharco, M. Wortsman, R. Wightman, C. Gordon, N. Carlini, R. Taori, A. Dave, V. Shankar, H. Namkoong, J. Miller, H. Hajishirzi, A. Farhadi, L. Schmidt, Openclip, 2021. URL: https://doi.org/ 10.5281/zenodo.5143773. doi:10.5281/zenodo.5143773, if you use this software, please cite it as below. [17] M. Oquab, T. Darcet, T. Moutakanni, H. V. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, R. Howes, P.-Y. Huang, H. Xu, V. Sharma, S.-W. Li, W. Galuba, M. Rabbat, M. Assran, N. Ballas, G. Synnaeve, I. Misra, H. Jegou, J. Mairal, P. Labatut, A. Joulin, P. Bojanowski, Dinov2: Learning robust visual features without supervision, 2023. [18] J. Sun, Z. Shen, Y. Wang, H. Bao, X. Zhou, LoFTR: Detector-free local feature matching with transformers, CVPR (2021).