Automatic Identification of Individual African Leopards in Unlabeled Camera Trap Images

This article describes an algorithm to solve the real-world animal identification problem, i.e., determine the unknown number of <tex-math notation="LaTeX">K</tex-math> individual animals in a dataset of <tex-math notation="LaTeX">N</tex-math> unlabeled camera-trap images of African leopards, provided by Panthera. To determine the leopards' IDs, we propose an effective automated algorithm, that consists of segmenting leopard bodies from images, scoring similarity between image pairs, and clustering followed by verification. To perform clustering, we employ a modified ternary search that uses a novel adaptive <tex-math notation="LaTeX">k</tex-math> -medoids <tex-math notation="LaTeX">$++$</tex-math> clustering algorithm. The best clustering is determined using an expanded definition of the silhouette score. A new post-clustering verification procedure is used to further improve the quality of a clustering. The algorithm was evaluated using the Panthera dataset that consists of 677 individual leopards taken from 1555 images, and resulted in a clustering with an adjusted mutual information score of 0.958 as compared to 0.864 using a baseline <tex-math notation="LaTeX">k</tex-math> -medoids <tex-math notation="LaTeX">$++$</tex-math> clustering algorithm. Note to Practitioners —We proposed an effective automated algorithm to solve the real-world animal identification problem: identifying <tex-math notation="LaTeX">K</tex-math> unknown individual animals in <tex-math notation="LaTeX">N</tex-math> images of a given species, with most animals only represented by a single image. This algorithm is different from other methods that assume all images in a dataset are from known individuals and thus regard the animal ID problem as a retrieval identification task. Our approach consists of a new adaptive <tex-math notation="LaTeX">k</tex-math> -medoids <tex-math notation="LaTeX">$++$</tex-math> clustering algorithm and a novel post-clustering verification procedure. The clustering is performed based on the degree of similarity between all image pairs in the dataset with the result validated using an expanded definition of the silhouette score. The accuracy of our algorithm was demonstrated on a real-world image dataset of African leopards, a small dataset with a relatively large ratio of <tex-math notation="LaTeX">$K/N$</tex-math> , provided by Panthera. Code has been made available at: https://github.com/obaiga/Automatic-individual-animal-identification.