High-Resolution with Global Context Network for Human Pose Estimation
Kehao Wang, Chenglin Li, Ruiqi Ren · 2022 27th Asia Pacific Conference on Communications (APCC) · 2022
Despite significant improvement in human pose estimation research, most top-performance methods are challenging to deploy in practical applications because of their complex architecture and high computational costs. Although the lightweight human pose estimation approach requires less processing and may be deployed on devices with low resources, such as mobile phones or robots, its network model performance is not exceptional. In this paper, we design the structure based on High-Resolution Network (HRNet), and propose a High-Resolution and Global Context Network (HRGCNet) based on the attention mechanism. Our approach redesigns the bottleneck block according to the attention mechanism of the Global Context Network (GCNet). By combining lightweight and high-performance GC blocks with bottleneck blocks, HRGCNet adds global context features at each location in the high-resolution subnet. The resulting high-resolution representation contains richer feature information. Our experiments on the COCO train2017 dataset show the efficiency of our method. Compared to HRNet with state-of-the-art performance, HRGCNet achieves higher accuracy, and the AP score improves by 2.0 percentage points with similar model size (#Params) and computational complexity (FLOPs). On the COCO test-dev set, HRGCNet has an AP score of 78.3, which is better than most current methods with good performance.