Fig. 3
From: Origin centric and part based pose decomposition for 3D human pose estimation
Architecture and computational flow of the origin-centric part transformer block. The input 2D skeleton feature undergoes two key processes: skeleton separation and skeleton recombination, resulting in the generation of OPart features. After each OPart feature is projected into a higher-dimensional space, spatial dependencies are modeled using scaled dot-product attention. Since the origin features span multiple OParts and carry positional dependencies with each part, they are converged into a single feature through a linear transformation module. Finally, the complete skeleton feature is restored via skeleton reconstruction.
