Lightweight Pose Estimation Based on Adaptive Robust Loss and Attention Mechanism
Lianxi Zhou · 2024
With the advancement of deep learning techniques, human pose estimation has become a key research area in computer vision. This study enhances the lightweight pose estimation framework, Lite-HRNet, by introducing Global Grouped Coordinate Attention (GGCA) and Multi-Scale Dilated Fusion Attention (MDFA) to improve the model's ability to capture both fine details and contextual information. Additionally, an adaptive robust loss function is employed to optimize the training process, strengthening the model's capability to handle outliers and improving overall robustness. Experiments on the COCO val2017 dataset show a 1.4% increase in mean Average Precision (mAP), while maintaining the same parameter count and computational complexity (GFLOPs), demonstrating the effectiveness of the proposed approach in boosting human pose estimation performance.