Pose-DWT-Former: An Improved Transformer-Based 3D Human Pose Estimation Model

Human pose estimation (HPE) is the basis of a wide variety of computer vision tasks. However, existing approaches are designed mainly for static 2D images, ignoring the temporal continuity and geometric consistency between video frames. With the aim of resolving the aforementioned issues, Pose-DWT-Former, an enhanced transformer-based approach, was proposed in this paper. The principle is to employ the 2D skeleton-based pose sequences extracted from the video frames and the discrete wavelet domain (DWT) information of those sequences as the network input for the purposes of 3D HPE. The evaluations were performed on two datasets, with good results in speed, accuracy, and robustness against noise, laying a solid foundation for the subsequent use of 3D human posture information for action recognition.