Mitigating retargeting-induced biases for atypical anthropometrics and pathological gait via large-scale physics-based training

Determine whether large-scale training in a physics-based dynamic simulator can mitigate inaccuracies introduced when retargeting SMPL-based motion capture to the MyoFullBody musculoskeletal morphology, particularly for individuals with atypical anthropometrics, asymmetric gait patterns, or musculoskeletal pathologies.

Background

The retargeting pipeline maps SMPL-based motion capture to the musculoskeletal models, but SMPL joints are statistical estimates derived from healthy populations. The authors note that this can misrepresent atypical anthropometrics or pathological conditions, causing discrepancies in joint centers, segment lengths, and moment arms that may smooth out or obscure clinically relevant features after retargeting.

They explicitly identify uncertainty about whether extensive training within a physics-based simulator could compensate for these retargeting shortcomings, posing it as an open question.