Origin centric and part based pose decomposition for 3D human pose estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Transformer models show promise in 3D human pose estimation from 2D images.
• Current methods struggle to balance global skeleton context with local joint features.

Purpose of the Study:

• To propose a novel transformer block, Origin-centric Part Transformer (OPFormer), for improved 3D human pose estimation.
• To address limitations in capturing both global and local spatial dependencies in 2D human skeletons.

Main Methods:

• Introduced the OPFormer block with Skeleton Separation and Skeleton Recombination steps.
• Skeleton Separation extracts fine-grained local spatial features.
• Skeleton Recombination uses a human skeleton Origin to fuse local and global features via a Spatial Transformer Encoder.

Main Results:

• Achieved state-of-the-art performance on Human3.6M and MPI-INF-3DHP datasets.
• OPFormer attained a Mean Per Joint Position Error (MPJPE) of 37.6mm on Human3.6M without extra training data.
• Demonstrated superior accuracy in 3D human pose estimation.

Conclusions:

• The OPFormer block effectively integrates local and global features for accurate 3D human pose estimation.
• This novel approach advances the field of 2D-to-3D human pose reconstruction.
• OPFormer offers a robust solution for precise human pose estimation from monocular 2D inputs.