We present a solution to the CLEF 2025 [1] animal species identification task, leveraging pretrained embedding models and a boosting classifier for pairwise similarity learning. Our pipeline combines pretrained feature extraction, pairwise embedding comparison, and supervised boosting to determine species-level similarity between image pairs. As a team "Tim Riggins" we have achieved competitive performance, obtaining scores of 0.618 target metric for our selected submission and 0.629 target metric for our best submission on the CLEF private leaderboard.
Metric learning aims to project images into an embedding space where semantically similar items are
close and dissimilar items are far apart. Common approaches include triplet loss [
Gradient boosting methods are widely used for structured and tabular data due to their strong
performance, robustness to overfitting, and ability to handle heterogeneous feature types. These models
iteratively build an ensemble of weak learners, typically decision trees, to minimize a loss function
through stage-wise additive modeling. Among popular implementations, LightGBM [
Open Set Recognition (OSR) aims to address the realistic scenario where a model may encounter
inputs from classes not seen during training. Unlike traditional closed-set classifiers, OSR models must
detect and reject unfamiliar samples rather than misclassify them. Techniques such as OpenMax [
Energy-Based Models (EBMs) provide a principled framework for modeling uncertainty and detecting
out-of-distribution inputs. Instead of outputting class probabilities directly, these models compute
an energy score that reflects the compatibility of input and model. In the OSR context, inputs with
high energy are flagged as unfamiliar or anomalous. Liu et al. [
The CLEF 2025 dataset includes images of individual animals across several species, such as lynxes,
salamanders, and sea turtles with the distribution shown in the Table 1, supported along with rich
metadata (e.g., species ID, individual ID, timestamp, and orientation) [
The goal of the entity matching task is to determine whether a query animal image belongs to a known
individual from a reference database or represents a previously unseen individual. The challenge spans
multiple species and requires models to handle significant visual variability while generalizing to novel
instances. Evaluation is based on two core metrics [
To preserve the aspect ratio of the original images, we applied padding before resizing. Images were then resized to 384 × 384 pixels when using the MegaDescriptor model, and to 512 × 512 pixels for all other local descriptor extractors. This preprocessing ensured consistency across the input pipeline while maintaining the structural integrity of key visual features.
For normalization, we used ImageNet statistics (mean and standard deviation) when working with MegaDescriptor. For the other models, min-max normalization.
MegaDescriptor [
We experimented with local descriptor extractors to match animal instances based on fine-grained
visual details. Specifically, we used three pre-trained models: ALIKED [
To combine keypoint information from multiple local descriptor extractors, we used LightGlue [
For each image pair, we compute cosine similarity and the number of corresponding points from local feature extractors Table 2 ending up with 9 features per sample pair. These serve as input features for the boost classifier.
To improve the robustness of local feature-based similarity, we applied a top- filtering strategy using the MegaDescriptor model. For each pair of images, we extracted local descriptors and retained only the most similar keypoint correspondences based on the distance of the descriptor. This step helps to reduce the noise from irrelevant or weak matches and emphasizes the most confident local alignments between images.
Utilizing cross-validation scheme, a smaller value of = 40 was suficient for lynxes and turtles, while salamanders required a higher threshold of = 150 to maintain performance, likely due to greater visual similarity and more challenging matching conditions within that class.
We train a LightGBM binary classifier to predict whether a given pair of images represents the same entities. Positive pairs consist of the same entities; negative pairs consist of diferent entities. To better simulate the conditions of the test set, we leveraged the timestamp information available in the Kaggle dataset to construct train–test splits that follow the natural chronological order of image collection.
In addition, we held out 20% of the individuals during training and used them as unseen entities. This setup allowed us to evaluate both identification of known individuals and detection of new ones, closely aligning with the evaluation protocol of the challenge.
Based on our validation results, we selected a threshold of 0.65 to distinguish new individuals from known ones.
During inference, we followed the same procedure as in training. For each query image, we identiefid its most similar reference images based on embedding similarity. These top- pairs were then passed through the boosting model to obtain binary match scores. The final prediction was assigned based on the reference image with the highest predicted similarity score among the candidates. Approach depicted in the Figure 2 allowed us to eficiently combine retrieval-based filtering with supervised matching for robust entity recognition.
We report performance in the Tables 4, 5 using the oficial CLEF evaluation metrics. Our model outperforms simple cosine thresholding and benefits from the added discrimination of the supervised boosting model. Both the use of boosting instead of averaging, and the application of padding, lead to improved results.
Despite the overall consistency between validation and private leaderboard scores, some discrepancies were observed. For instance, the configuration with a threshold of 0.75 achieved the highest private score (0.629), while a lower threshold of 0.6 yielded the best validation score (0.686). This suggests that our validation split, although timeline-based and stratified, may not perfectly reflect the distribution or dificulty of the private test set. Overfitting to the validation threshold likely accounts for this performance gap, highlighting the importance of evaluating across multiple thresholds and data partitions. In future work, incorporating more robust cross-validation or ensembling across decision thresholds could help bridge this mismatch.
Finetuning MegaDescriptor within a multi-class classification framework led to noticeable performance improvements over using frozen embeddings. The model was trained to predict individual entities directly, which encouraged more discriminative representations. However, this approach proved computationally expensive and required extensive training data for convergence. In contrast, keypoint-based matching methods ofered a more eficient alternative, especially when used with pretrained extractors and lightweight post-processing. As a result, we prioritized methods that combined pretrained descriptors with retrieval and pairwise scoring.
This pipeline demonstrated strong generalization performance and remained stable across validation and test scenarios. By combining global and local features with a lightweight boosting classifier, it efectively captured both coarse and fine-grained visual similarities. The method required minimal ifnetuning and was computationally eficient at inference time. As a result, it secured 7th place on the private leaderboard of the CLEF 2025 animal re-identification challenge.
We evaluated several baseline strategies to understand the contribution of each component in our pipeline. A simple cosine similarity between pretrained embeddings yielded limited performance, especially in distinguishing hard negative pairs. We also experimented with training classification models and using their penultimate-layer embeddings for cosine-based retrieval; while this improved over the vanilla approach, it still lacked robustness. In contrast, our boosting-based method, which leverages pairwise features, consistently outperformed both baselines by capturing more nuanced relationships between image pairs. This highlights the value of supervised modeling on relational features beyond raw embedding distances.
The pretrained embedding model captured useful semantic structure. Boosting provided flexible decision boundaries in embedding space. Triplet-based training was unstable in our setup, and the pretrained matcher consistently outperformed self-trained models.
We presented a simple yet efective framework for entity matching based on pretrained visual embeddings and a supervised boosting model trained on pairwise features. Rather than relying solely on embedding distance thresholds, our approach learns to model similarity through handcrafted relational features and a gradient boosting classifier, providing more flexibility and robustness to noise or domain shifts. This framework proved particularly useful in scenarios where the embedding space alone was insuficient to capture fine-grained species-level distinctions. In future work, we plan to explore selfsupervised pretraining to improve the embedding quality, incorporate graph-based label propagation to exploit the structure of embedding similarity graphs, and investigate hierarchical clustering techniques to capture taxonomic relationships between species.
During the preparation of this work, the author(s) used ChatGPT in order to: Grammar and spelling check. The changes were minimal and afected only grammar and punctuation, not the meaning of the text. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.
We thank the CLEF 2025 organizers and baseline authors for providing pretrained models and evaluation scripts.
All code can be found at the github repository https://github.com/SergeyFedorchenko/AnimalCLEF25_ 7th.git.