3DMDPose: 3D Human Pose Estimation based on Multi-Level Features and Diffusion Modeling

D human pose estimation suffers from selfobscuration, 2D-to-3D pose mismatch, and computational complexity. 3D HPE based on complex environments is very challenging. Existing pose estimation in self-attention blocks is computational complexity. Moreover, its complexity grows quadratically with the number of input sequences. We explore a new 3D pose estimation block called MLN, which use a multilayer perceptron with normalization layer. This block shows better results in the absence of self-attention, allowing the model to better handle pose estimation in complex scenes. Meanwhile, toward solving the problem that the MLP model can't capture the local details well, which makes the model lack the a priori knowledge of human body configurations. We proposed another refinement module, Multi-level feature processing (MFNet). The module processes and integrates joint information at different levels to optimize the pose estimation results at both local and global levels. We propose a new framework that integrates MLN and MFNet, named 3DMDPose. Extensive experiments were conducted on the Human3.6M and MPI-INF3DHP datasets. Our framework achieved the most favorable performance results in 3DHPE.