| Title |
Rethinking Rotation Representation for Human Mesh Recovery from Vertex Sequences |
| Authors |
전성호(Sungho Chun) ; 장주용(Ju Yong Chang) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.6.91 |
| Keywords |
Absolute rotation; Relative rotation; Mesh vertices; SMPL parameter |
| Abstract |
This paper addresses the vertices-to-parameters problem of inferring SMPL parameters from a 3D human mesh vertex sequence. This process is essential for converting high-resolution 3D scan data into a controllable parametric model or for refining meshes corrupted by noise. Existing human mesh reconstruction methods have conventionally inferred pose parameters in the form of relative rotations with respect to parent joints. However, given that the input mesh vertices inherently contain absolute positional information in a global coordinate system, predicting relative rotations imposes an additional burden on the network to implicitly learn the complex human kinematic chain. Based on the hypothesis that aligning the coordinate systems of the input global vertices and the output rotations can reduce the complexity of the mapping function and improve learning performance, we propose to directly infer absolute rotations instead of relative rotations. To validate this idea, we design a transformer encoder-based Seq2Seq model optimized for spatio-temporal feature extraction, and conduct comparative experiments under an identical architecture by varying only the representation of the output head between relative and absolute rotations. Experimental results on the AMASS dataset demonstrate that the absolute-rotation-based model consistently achieves superior accuracy and stability for both ideal meshes without noise and realistic meshes containing noise. These results indicate that, in deep learning-based inverse kinematics problems, the choice of a parameter space that is well aligned with the data characteristics plays a decisive role in performance. |