MPE-HRNet: A Lightweight High-Resolution Network for Multispecies Animal Pose Estimation

Area of Science:

• Computer Vision
• Machine Learning
• Animal Behavior Analysis

Background:

• Deep learning is increasingly vital for animal pose estimation in health, conservation, and behavior studies.
• Deploying pose estimation models on resource-constrained edge devices necessitates balancing model complexity and accuracy.

Purpose of the Study:

• To propose MPE-HRNet.L, an improved lightweight network model based on Lite-HRNet for efficient animal pose estimation.
• To enhance the performance of animal pose estimation models for deployment on edge devices.

Main Methods:

• The MPE-HRNet.L model was developed by enhancing Lite-HRNet with improved Spatial Pyramid Pooling-Fast in different branches.
• A novel feature extraction module incorporating mixed pooling and dual attention mechanisms was designed as the core component.
• An additional feature enhancement stage was introduced to refine critical features within the network.

Main Results:

• Experimental validation on the AP-10K and Animal Pose datasets demonstrated the effectiveness of MPE-HRNet.L.
• The proposed model achieved a favorable balance between accuracy and computational efficiency.

Conclusions:

• MPE-HRNet.L offers an effective and efficient solution for animal pose estimation on edge devices.
• The model's design innovations contribute to advancing deep learning applications in animal behavior and welfare.