Abstract
Accurate parametric 3D human body models are relevant for several applications, including medical use cases such as burn care, where patient-specific body models can
support the estimation of the percentage of burned body surface area which is a key
metric for burn assessment and treatment planning. Generating these models from sim
ple anthropometric measurements provides a cheaper and more practical alternative to
3D body scanners and image-based reconstruction methods, especially for patients with
limited mobility. However, anthropometric measurements are affected by measurement
uncertainty and differ systematically from measurements derived from parametric body
models. These inconsistencies can bias the reconstruction process and may result in
body shapes that match the input measurements but are anatomically implausible.
This thesis investigates the reconstruction of SMPL body shapes from anthropo
metric measurements, with a focus on estimating accurate and anatomically plausible
SMPL shape parameters from measurements obtained under real-world measurement
conditions. For this purpose, multiple SMPL-based and anthropometric datasets with
different acquisition modalities are used, including manually measured, scan-derived,
and SMPL-derived measurements. A dataset containing manual tape measurements and
corresponding fitted SMPL body models for the same subjects is used for mesh-level
evaluation.
Several regression models and a per-sample optimization method are investigated. In
addition to standard regression models, a custom loss function is introduced to optimize
the reconstructed anthropometric measurements while regularizing the estimated SMPL
shape parameters. Among the evaluated approaches, the per-sample optimizer offers the
greatest practical flexibility, as it can use any available subset of measurements and does
not require missing measurements to be explicitly imputed.
The results show that 3D body models can be reconstructed from a limited set of
anthropometric measurements with reasonable accuracy. In the final evaluation, the
per-sample optimizer achieves a mean per-vertex error of 12–17 mm. The generated
models show reasonable global shape accuracy, although larger local deviations remain
in regions that are not sufficiently constrained by the available measurements. Overall,
this thesis demonstrates that anthropometric measurements can serve as a promising
input modality for generating patient-specific 3D body models, while showing that mea
surement uncertainty and modality-dependent differences must be carefully considered.
| Date of Award | 2026 |
|---|---|
| Original language | English (American) |
| Awarding Institution |
|
| Supervisor | Ulrich Bodenhofer (Supervisor) |
Studyprogram
- Data Science and Engineering
Cite this
- Standard