AJA-Pose: A Framework for Animal Pose Estimation Based on VHR Network Architecture

Pose estimation is the process of identifying key points on an animal's body, and is critical for understanding animal behaviour, movement, and interactions within their natural habitats. Despite its importance, pose estimation in animals, especially across a broad range of species, poses unique challenges due to the diversity in anatomical structures, sizes, and the complexity of natural environments. This work uses the Animal Kingdom dataset to fine-tune the VHR-BirdPose: Vision Transformer-Based HRNet for Bird Pose Estimation with Attention Mechanism, resulting in AJA-pose a PyPi package framework for training, and evaluating animal pose estimation. The package allows increased access to the VHR-BirdPose model enabling reproducibility and experimentation. The fine-tuned model achieved an accuracy of 93.07% under the validation set. We show that our model outperforms the earlier publication trained on only birds, 90.19%, with ours trained on all animals, 95.16%. This paper describes our participation in the 2024 ICME Grand Challenge under the Pose Estimation Category. The objective of the challenge was to predict the keypoints of 850 animal species.