Deterministic Measurement-Conditioned 3D Body Reconstruction for Precision Digital Fitting

Abstract

Custom tailoring and digital fitting require geometrically precise avatars, while most existing methods prioritize visual realism. We present a method for reconstructing three-dimensional human bodies from 47 anthropometric measurements, with accuracy within one standard clothing-size step, ensuring suitability for pattern construction. We train a separate model per sex, each on 400 real scans, using a deterministic FiLM-conditioned regression network with per-dimension Fourier feature mapping, pose-robust geometric losses, and scale augmentation. We show that dense anthropometric measurements carry most of the shape information relevant to fitting. In our tests the deterministic model reaches lower measurement error than a conditional variational autoencoder and, unlike an unconstrained linear baseline, stays geometrically stable on out-of-distribution measurements.