Abstract
In recent years, many Vision Transformers (ViTs)-based methods have become popular in the field of Human Pose Estimation (HPE) and have achieved excellent results. However, Convolutional Neural Networks (CNNs)-based architectures still have many advantages in the field of HPE. Therefore, we have combined the strengths of the ViTs framework and applied it to CNN-based HPE methods. We introduce a novel 2D HPE method called the PoseAlign Network for Hybrid Structure (PAN-HS). PAN-HS leverages the conceptually simple yet effective depth-wise convolution to design two feature extraction blocks: the Spatial Align Block and the Channel Align Block. These blocks improve the representation of mid-level spatial and channel features. By efficiently aggregating features of the same spatial size (intra-layer features), PAN-HS achieves a fine-grained localized representation that preserves rich spatial information, thereby improving the precision of keypoint localization. Additionally, we observe varying contributions of different output features to the final performance. We introduce an efficient attention mechanism called Point Reposition Attention, which balances the local and global representations of output features and optimizes keypoint location predictions through rescaling. Our method achieves a mean PCKh of 92.74% on the MPII Dataset, demonstrating outstanding performance and superior pose estimation capabilities.