A Unified Approach to Pose Estimation in Elephants and Other Quadrupeds using Noisy Labels

Pose estimation predicts anatomical landmarks in humans and animals from monocular images or videos. Animal Pose Estimation is crucial for monitoring locomotion, behavior, and activity recognition, playing a key role in wildlife conservation. Single species pose estimation studies capture features unique to the species but generalize sub-optimally, while multi-species studies provide broader generalization by assuming fixed keypoints for all quadrupeds, this oversimplification fails to capture unique anatomical traits in animals such as elephants. To harness the strengths of single-species and multi-species pose estimation, we present QuadPose, a framework that standardizes skeletal structures across datasets and improves generalizability through consistency-dependent pseudo-labelling. Additionally, JumboPose, a manually annotated dataset of 2,078 African elephant images with 33 keypoints tailored to their unique morphology is introduced. Extensive evaluations demonstrate the effectiveness of QuadPose for animal pose estimation. This work establishes a foundation for standardized, cross-species pose estimation, advancing applications in wildlife conservation, and veterinary research.