ScholarWorks@숙명여자대학교

초록

This paper presents a multi-matcher fusion pipeline developed for the AnimalCLEF 2025 individual animal re-identification challenge. The task involves identifying distinct individuals within the same species, under constraints such as one-shot learning and open-set recognition for previously unknown individuals. To address these challenges, we integrate three complementary matchers: MegaDescriptor for global visual features, ALIKED for local keypoint-based matching, and EVA02 for semantic-level similarity. These components are fused using WildFusion-based score calibration and a simple weighted averaging scheme. Additionally, species-specific preprocessing, such as orientation normalization for salamanders and 5-crop Test-Time Augmentation, is applied to enhance robustness. Our final pipeline achieved a public score of 0.50708 and a private score of 0.53185, representing a 23.2 percentage points relative improvement over a baseline solution (0.3002). According to the official leaderboard, our system ranks 44th out of 230 participating teams, placing in the top 19%. This outcome demonstrates the effectiveness of combining global, local, and semantic descriptors through calibrated fusion in a multi-species wildlife ReID context. The full implementation of our pipeline is available at https://github.com/dongyeon1031/AnimalCLEF2025. © 2025 Copyright for this paper by its authors.

키워드