An efficient and accurate 2D human pose estimation method using VTTransPose network

Abstract Human pose estimation is a crucial area of study in computer vision. Transformer-based pose estimation algorithms have gained popularity for their excellent performance and relatively compact parameterization. However, these algorithms often face challenges including high computational demands and insensitivity to local details. To address these problems, the Twin attention module was introduced in TransPose to improve model efficiency and reduce resource consumption. Additionally, to address issues related to insufficient joint feature representation and poor network recognition performance, the enhanced TransPose model, named VTTransPose, replaced the basic block in the third subnet with the intra-level feature fusion module V block. The performance of the proposed VTTransPose model was validated on the public datasets COCO val2017 and COCO test-dev2017. The experimental results on COCO val2017 and COCO test-dev2017 indicate that the AP evaluation index scores of the VTTransPose network proposed are 76.5 and 73.6 respectively, marking improvements of 0.4 and 0.2 over the original TransPose network. Additionally, VTTransPose exhibited a reduction of 4.8G FLOPs, 2M parameters, and approximately 40% lower memory usage during training compared to the original TransPose model. All the experimental results demonstrate that the proposed VTTransPose is more accurate, efficient, and lightweight compared to the original TransPose model.