The AnimalCLEF2025 task aims to train a recognizer on the training set with good generalization ability for predicting whether the category of an animal image in the test dataset belongs to a known class or an unknown class, which is an open-set recognition problem. In this study, the fusion of deep features and low-level features is utilized to build an open-set animal recognition model. Specifically, the pretraining and fine-tuning scheme is adopted to learn high-level features, where the swin-base-patch4-window7-224 model is fine-tuned with the training set and deep features are then extracted from its feature representation layer. Second, we utilize two keypoint detection and descriptor extraction networks to obtain two sets of low-level features. Afterwards, the nearest-neighbor classification rule is utilized with the weighted multilevel features to infer the label of a test sample. Particularly, if the maximal similarity between the test sample and training samples is lower than a threshold, the test sample is predicted as unknown classes. Finally, experimental results show that the proposed model obtains 0.55826 and 0.56044 balance accuracy on the 31% and 69% test dataset, respectively. Our source code is available at https://github.com/NickyTan8899/tjy.
Open-set recognition (OSR) refers to the problem where a model is trained on a set of known classes
(the closed set), but during testing, it should correctly classify known categories and identify instances
from unknown classes as unknown. This setting better reflects real-world applications than traditional
closed-set recognition that assumes that all test samples belong to the known classes[
Accordingly, researchers have explored and designed a variety of models towards enhanced open-set
recognition accuracy. Unlike traditional closed-set classifiers, OSR methods incorporate mechanisms
to detect novel instances. Approaches to OSR can be broadly categorized into discriminative models,
generative models, and distance-based methods[
For the problem of open-set animal recognition, though great progresses have been made, most
of existing methods still sufer from degraded performance due to complex spatial dependencies and
background information[
(1) Deep features and low-level features are extracted. Particularly, deep features are learnt under the pretraining and fine-tuning scheme, where the swin-base-patch4-window7-224 model is utilized and ifne-tuned. Keypoint detection and descriptor extraction networks are also used to extract features.
(2) The open-set recognition is performed with the multilevel features. The similarity between a test sample with training samples are measured. Specifically, we first calculate the similarity from the view of deep features and low-level features respectively, and then use the Wildfusion strategy to obtain the final similarity score[ 9]. The prediction is made according to the similarity score. If the maximal similarity is lower than a threshold, we categorize it into novel classes; otherwise, we report the class of training sample corresponding to the maximal similarity.
(3) Comparative experiments are conducted against two baseline models on the competition datasets. Results show that the proposed model performs better than the baselines and obtains 0.55826 and 0.56044 balance accuracy on 31% and 69% test dataset, respectively, with the best result ranked 34th on the leaderboard (Team Name: Already mygo).
The structure of this paper is as follows. Section 2 introduces the proposed open-set animal recognition model. Section 3 presents experimental datasets and experimental setup. Section 4 presents experimental results, followed by the conclusion section.
As for deep features, we adopt the pretrain and fine-tuning scheme to learn latent features. Specifically, a Swin Transformer-based pretrained model (i.e., swin-base-patch4-window7-224 model)[10] is adopted. We replace its cross-entropy loss with ArcFace loss that enhances the angular separation between diferent classes in the embedding space[ 11]. Afterwards, we fine-tune the pretrained model on the training set. Finally, we drop its classification layer to obtain a feature learning network, which enables us to obtain the feature representations of an image. For clarity, we denote the set of features as fs1.
Besides, we directly apply two diferent keypoint and descriptor extraction networks (i.e., ALIKED and DISK) on the image to obtain two sets of features. Specifically, ALIKED[ 12] utilizes the sparse deformable descriptor head to extract deformable descriptors, where a neural reprojection error loss is used to measure the discrepancy between reprojection and descriptor-matching probabilities[13]. DISK[14] uses reinforcement learning to train end-to-end pipeline for keypoint detection and matching. We denote the two sets of features as fs2 and fs3.
In classifying a test sample, we adopt the nearest-neighbor classification rule to make predictions. First, we encode the test sample and each of the training samples with fs1, fs2, and fs3. Second, we measure the similarity between test sample and each training sample. Particularly, to reflect the importance of the three diferent types of features, the Wildfusion strategy is adopted, with which we can get the similarity between test sample and each training sample and further obtain the maximal similarity. If the maximal similarity is greater than a predefined threshold, the predicted result of the test sample is the label of the associated training sample; otherwise, the predicted label is “unknown classes” (or new_individual used in the competition).
The AnimalCLEF2025 dataset, having 15209 images, is divided into database dataset (having 13074 images) and query dataset (having 2135 images). The database plays the role of training set and the query serves as the test set. Table 1 presents the summary of experimental data.
To obtain deep features, we first resize the images to 224 x 224 to make it suitable for the pretrained Swin-based model. We then fine-tune it on the database dataset with the SGD optimizer. An initial learning rate 0.001 is used and a cosine annealing learning rate scheduler is utilized, which adjusts the learning rate following a cosine curve from the initial value decreased to 1e-6 in the end. The model is trained for 100 epochs. Meanwhile, the batch size and the number of workers for loading data are 64 and 2 respectively. Finally, the deep features are embedded to be a 1 x 1024 vector for each sample. As for low-level features, we resize the images to 256 x 256 and then obtain two sets of features, returned by ALIKED and DISK, respectively.
Besides, for comparison, three baseline methods including MegaDescriptor-L-384 (swin-large-patch4window12-384 model), MegaDescriptor-L-384-ALIKED (a combination of MegaDescriptor-L-384 and ALIKED) and MegaDescriptor-L-384-ALIKED-DISK (a combination of MegaDescriptor-L-384, ALIKED and DISK) are utilized to demonstrate the efectiveness of our proposed model.
As for the performance metric, balance accuracy (denoted by score in formula (1)) is used, which is calculated by the balanced accuracy on known samples (BAKS) and balanced accuracy on unknown samples (BAUS).
score = √BAKS × BAUS. (1) , where BAKS is the accuracy of individuals that are known in the database, and BAUS is the accuracy of individuals that are unknown in the database. This formulation avoids misleadingly high scores from trivial models, such as those predicting all samples as unknown, which would score 0% BAKS and 100% BAUS. Unlike the arithmetic mean, the geometric mean penalizes such imbalance, providing a more robust metric.
Furthermore, to investigate the impact of diferent thresholds on the recognition performance, we conduct experiments with the candidate thresholds ranging from 0.3 to 0.8. Figure 2 presents the results,
model
MegaDescriptor-L-384
MegaDescriptor-L-384-ALIKED MegaDescriptor-L-384-ALIKED-DISK
Ours score from which we can observe a general trend that the score first increases and then decreases along with the increase of the threshold. We also observe that the use of 0.4 generally leads to better performance.
Towards higher image-based open-set animal recognition, we in this study propose a distance-based recognition model that utilizes several sets of features. The deep features are learnt with the pretraining and fine-tuning scheme and the low-level features are obtained by the keypoint detection and descriptor extraction networks. To reflect the importance of diferent types of features in calculating the similarity between samples, a weighing scheme is used. Afterwards, a threshold-based method is adopted to infer the label of a test sample. Finally, the proposed model is evaluated on the test set and results demonstrate its efectiveness.
During the preparation of this work, the authors used OpenAI-GPT-4o in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [9] V. Cermak, L. Picek, L. Adam, L. Neumann, J. Matas, Wildfusion: Individual animal identification with calibrated similarity fusion, in: European Conference on Computer Vision, Springer, 2025, pp. 18–36. [10] Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, B. Guo, Swin transformer: Hierarchical vision transformer using shifted windows, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10012–10022. [11] J. Deng, J. Guo, N. Xue, 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. [12] X. Zhao, X. Wu, W. Chen, P. C. Chen, Q. Xu, Z. Li, Aliked: A lighter keypoint and descriptor extraction network via deformable transformation, IEEE Transactions on Instrumentation and Measurement 72 (2023) 1–16. [13] H. Germain, V. Lepetit, G. Bourmaud, Neural reprojection error: Merging feature learning and camera pose estimation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 414–423. [14] M. Tyszkiewicz, P. Fua, E. Trulls, Disk: Learning local features with policy gradient, Advances in Neural Information Processing Systems 33 (2020) 14254–14265.