Spot the Difference: An Image-Based Approach for Classifying Levels of Relatedness of Giraffes

Abstract

Image-based animal identification provides a more cost-effective, less invasive avenue for monitoring populations, including giraffes. This becomes increasingly important as giraffe populations continue to experience significant threats resulting in population declines from historic numbers. Yet, leading giraffe identification software, GiraffeSpotter and underlying model MiewID-v3, is limited to identifying the same individual based on the right side coat pattern. This project works to extend the MiewID-v3 model to classify other levels of relatedness based on their coat patterns. This project contributes (1) two novel datasets of Reticulated Giraffes, one with images of both the right and left sides of individuals, and one with right images of giraffes and corresponding information on which individuals were sighted in the same encounter group, (2) the first image-based model that effectively classifies giraffes by species, (3) a demonstration that the general animal identification featurization in the MiewID-v3 model can capture some of the similarity of individuals' coat patterns at different levels of relatedness, (4) a definition of a classification task to match giraffes based on their coat patterns at various levels of relatedness, (5) an initial comprehensive models for developing a feature space where the distances between embedding vectors correspond to different levels of relatedness. This project hopes to contribute to understanding the relationships between individual giraffes for conservation applications.