SOMA: From Surface Observations to Muscle Anatomy

How to read this

Category

Person-specific inverse model recovering spatio-temporal muscle and skin deformation from multiview RGB video

Contributions

• Poses a novel inverse research problem, recovering muscle deformation directly from visible skin surface observations and pose, rather than only simulating muscle activation forward with FEM or biomechanical tools
• Presents SOMA, a person-specific model using cascaded U-Nets that infers spatio-temporal muscle bulging and skin sliding from multiview RGB surface signals, described by the authors as the first method to attempt recovering muscle deformation from multi-view RGB data
• Introduces SKIM, a subject-specific soft-tissue deformation dataset, and produces anatomically grounded animation, including individual-muscle targeting, without the computational cost of full FEM or biomechanical simulation

Context

SOMA positions itself between two established but limited lines of work, parametric human body surface models (the SMPL family) that capture only skin geometry with no biomechanical insight, and biomechanical simulation tools that model muscle activation and force but do not connect those activations to observable external shape. It is explicitly the inverse of the usual biomechanics-to-shape direction.

Correctness

The claims rest on the purpose-built SKIM dataset, collected via RGB cameras for training and evaluation, with data and code stated as available for verification. The model is person-specific, so cross-subject generalization is outside its claims, and the claim of being first to recover muscle deformation from video should be read as the authors' own framing rather than something independently confirmed here.

Clarity

A clearly written ECCV-style paper with a crisp problem framing in the introduction (the inverse muscle-recovery problem), accessible to a character TD familiar with parametric body models even without deep biomechanics training.

How to read it

First pass: abstract and Figure 1's skin and muscle deformation outputs. Second pass: the introduction for the inverse-problem framing and how it differs from FEM and heuristic anatomy fitting. Third pass: the cascaded U-Net method section and the SKIM dataset description, plus the results for quantitative muscle-deformation accuracy.