Effective approaches for individual identification of African leopards in unlabeled camera trap images

Abstract

This dissertation proposes effective solutions to the real-world animal identification problem: identifying K unknown individual animals from N unlabeled camera-trap images of a given species, specifically African leopards. The research evolves from a fully automated algorithm to a human-in-the-loop framework, achieving high identification accuracy while substantially reducing human annotation effort. These approaches are particularly effective for small, unlabeled camera-trap image datasets of species that exhibit distinctive, identifiable markings and a high individual-to-image ratio, especially when many animals appear in only a single image. The fully automated identification algorithm begins by segmenting leopard bodies from images, computing pairwise image similarity scores, and performing clustering using a novel adaptive k-medoids++ clustering algorithm guided by a modified ternary search. The best clustering is determined through an expanded definition of the silhouette score, followed by a post-clustering verification procedure that further refines clustering quality. Evaluated on the Panthera dataset of African leopards, comprising 677 individual leopards captured in 1555 images, the algorithm achieves an adjusted mutual information score of 0.958, outperforming the baseline k-medoids++ algorithm. Building upon this foundation, the human-in-the-loop framework addresses the remaining challenges of disambiguating similarity scores between images of the same animal and those of different animals. This area is where fully automated algorithms remain difficult and thus necessitate limited human expert involvement. In this framework, a human expert is only required to determine whether a given image pair belongs to the same individual or not. Their input is incorporated across three stages: determining an appropriate initial positive threshold, confirming automatically detected potential positive image pairs, and verifying the consistency of a small, selectively chosen subset of autonomously identified positive image pairs. Despite this limited human participation, the algorithm maintains identification accuracy comparable to a human baseline method, in which each image is manually compared with its top-2 most similar images, while reducing human effort by 77.3%, requiring only 0.05% of all image pairs to be labeled. Additionally, an expanded and enhanced definition of the silhouette score is introduced to provide a more accurate evaluation of clustering performance. Overall, the proposed algorithms provide effective unsupervised and semi-supervised solutions for individual animal identification, offering particular value to researchers studying new habitats or species for which deep learning approaches are infeasible due to limited labeled data.