3D reconstruction of wild animals for re-identification

Abstract

Animal re-identification involves recognizing individual animals within the same species or interspecies by utilizing their distinct discriminative characteristics, such as fur patterns and other natural markings. It is essential for non-invasive monitoring of wildlife improving ecological research and conservation initiatives. Wildlife re-identification poses significant challenges compared to identifying objects in a controlled environments such as human re-identification in images and videos, due to variations in appearance, pose, and camera viewing angles. Solving this challenge may involve enhancing methods in animal re-identification, such as extracting 3D data to incorporate information about shape and pose, depth, and geometric structure into animal re-identification pipelines leading to model reliability and adaptability across different species. 3D reconstruction techniques have been applied in human re-identification (example, 3D face recognition), however, their application to animals remains limited. This thesis explores the gaps in animal 3D reconstruction and its incorporation into wildlife monitoring (re-identification). Two reconstruction paradigms were explored: an end-to-end (deep learning) learning-based approach and an optimization-based method that fits a template model to 2D observations. The result showed how optimization-based method outperformed some learning-based methods in accuracy while maintaining reasonably fast computation time. Experiments utilizing 3D texture maps reconstructed withWonder3D++ for Saimaa ringed seals demonstrated that 3D textures result in denser detected feature matching when compared to the original 2D input images, potentially improving the re-identification accuracy.