Skip to content

← ArchivePaper2025

Reconstructing Humans with a Biomechanically Accurate Skeleton

Yan Xia, Xiaowei Zhou, Etienne Vouga, Qixing Huang, Georgios Pavlakos

CVPRAcademicSkinningMusclesRetargeting

HSMR: first end-to-end image-to-SKEL regressor, jointly recovering the biomechanically accurate skeleton and body mesh from a single photograph.

Open the paper ↗Code ↗Trace in the lineage graph

Abstract

HSMR is the first end-to-end approach for reconstructing 3D humans from a single image by estimating the parameters of the biomechanically accurate SKEL model. A transformer is trained to estimate SKEL parameters from image inputs. Due to the lack of training data, the authors build a pipeline to produce pseudo ground truth model parameters and implement a training procedure that iteratively refines them. HSMR achieves competitive performance on standard benchmarks and significantly outperforms prior methods in settings with extreme 3D poses and viewpoints, while producing more realistic joint rotation estimates by respecting anatomical joint limits. Accepted as CVPR 2025 Oral.

How to read this

Category
Method: monocular human reconstruction with a biomechanical skeleton
Contributions
  • HSMR, the first end-to-end image-to-SKEL regressor recovering a biomechanically accurate skeleton and body mesh from a single image
  • A transformer that estimates SKEL parameters, trained via a pipeline that produces pseudo ground-truth and iteratively refines it
  • Stronger results on extreme poses and viewpoints with more realistic, anatomically constrained joint rotations
Context
Builds directly on the SKEL biomechanical body model (Keller et al. 2023), bringing it into the single-image human mesh recovery setting.Builds on: From Skin to Skeleton: Towards Biomechanically Accurate 3D Digital Humans
Correctness
Validated on standard benchmarks (competitive) plus extreme-pose and viewpoint settings (clear gains), but accuracy depends on iteratively refined pseudo ground-truth, so estimates inherit any bias in that bootstrapped supervision.
Clarity
An accepted oral, so the narrative is likely clear; a first pass gives the pipeline, a second pass covers the pseudo-GT refinement loop.
How to read it
Focus first on how SKEL parameters are regressed and how anatomical joint limits are enforced; second pass on the pseudo-ground-truth generation and refinement procedure if you plan to retrain or trust the joint angles.

Builds on

Built upon by

Nothing yet.

Related work

Keywords

human mesh recoverybiomechanically accurate skeletonskel modelsmplhsmrskelifypseudo ground truth refinementvit-h backbonevitposetransformer head6d rotation representationeuler anglesjoint limit violation2d keypoint reprojectionsmplifyhmr2.0body shape estimationvolume estimationmonocular pose estimationiterative fitting

This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →