OSSO: Obtaining Skeletal Shape from Outside

How to read this

Category

Method: predicting internal skeleton from body surface

Contributions

• OSSO, presented as the first method to learn the mapping from a 3D body surface to internal skeletal anatomy from real data
• A novel part-based 3D skeleton model paired with the STAR body model, learned from 2,000 DXA scans
• Public release of code and a paired skin/bone mesh dataset

Context

Builds on the STAR parametric body model (Osman et al. 2020), pairing its surface shape with a learned skeleton to recover inside-the-body anatomy.Builds on: STAR: Sparse Trained Articulated Human Body Regressor

Correctness

Trained on 2,000 DXA scans and constrained for physical plausibility, so predictions reflect that scan population and the imposed constraints; bone geometry is inferred from the outside rather than directly measured for arbitrary new subjects.

Clarity

Accessible and well-scoped; a first pass conveys the surface-to-skeleton goal, a second pass covers the part-based skeleton model.

How to read it

Focus on the data pairing and the part-based skeleton representation; a single careful pass suffices unless you plan to use the released dataset, then do a second pass on the training setup.