Parallel Convolution and Attention for Pose Estimation

Multi-scale feature fusion modules exhibit excellent performance in CNNs. However, these convolution modules suffer from inherent receptive field constraints for human pose estimation (HPE). Recently, transformer blocks have been introduced into HPE to exploit global information and capture correlations between different joints. But relying solely on the transformer block is insufficient and computationally expensive. We propose a parallel convolution and attention(PCA) module, which takes advantage of both local and global representations for efficient feature extraction. Furthermore, we propose a network based on the PCA module for HPE, termed PCApose. Experiments on MS-COCO demonstrate that, under comparable performance, PCApose reduces computation by half compared to Transpose [1]. Our model strikes a balance between computational burden and accuracy.