ZebraPoseNet: A Deep UNet-Based Framework with Attention Mechanisms for Animal Pose Estimation

Abstract

Animal pose estimation plays a crucial role in wildlife monitoring, behavioral analysis, and conservation research. However, zebras present unique challenges due to their visually complex striped patterns, frequent occlusions in natural environments, and the scarcity of annotated datasets. Conventional deep learning models, such as UNet, struggle with feature extraction and keypoint localization in such scenarios, leading to reduced accuracy and generalization issues. In this study, we propose an improved UNet model that incorporates Squeeze-and-Excitation (SE) attention mechanisms and transfer learning to enhance keypoint detection accuracy. The SE blocks allow the model to emphasize important spatial features while suppressing background noise, and the transfer learning approach leverages knowledge from larger animal pose datasets to improve performance on limited zebra data. We evaluate our model on a custom-labeled zebra dataset and optimize it with a hybrid loss function. Experimental results demonstrate that our approach significantly reduces the mean per-keypoint error (MPKE) by 15% compared to the baseline UNet model, highlighting its effectiveness in real-world applications.