animal re-identification

This study introduces a novel semi-supervised classification method that incorporates automatically generated pairwise constraints—Must Link (ML) and Cannot Link (CL)—from video data to improve animal re-identification accuracy. By applying these constraints frame-by-frame and combining them with standard classifiers, the proposed methods significantly outperform traditional classifiers, constrained clustering, and other semi-supervised learning approaches across five animal video datasets.

This study evaluates 25 classification models and five feature extraction techniques—including deep learning and traditional methods—for animal re-identification using annotated video datasets. Surprisingly, simple linear classifiers like Linear Discriminant Analysis (LDA) combined with basic RGB colour features outperformed deep learning models, highlighting the importance of comparative testing and the potential effectiveness of simpler approaches in complex, real-world scenarios.