A Neural Anthropometer Learning from Body Dimensions Computed on Human 3D Meshes

TL;DR

This paper introduces a method to accurately compute human body dimensions from 3D meshes and trains a neural network to estimate these measurements from images, addressing data scarcity and establishing a reproducible benchmark.

Contribution

It presents a novel approach to calculate key human body measurements from 3D meshes and develops a neural anthropometer trained on synthetic data for reliable measurement estimation.

Findings

Mean estimate error of 20.89 mm (2.84% relative error)

Provides a reproducible baseline for HBDE research

Enables fair comparison of future methods

Abstract

Human shape estimation has become increasingly important both theoretically and practically, for instance, in 3D mesh estimation, distance garment production and computational forensics, to mention just a few examples. As a further specialization, \emph{Human Body Dimensions Estimation} (HBDE) focuses on estimating human body measurements like shoulder width or chest circumference from images or 3D meshes usually using supervised learning approaches. The main obstacle in this context is the data scarcity problem, as collecting this ground truth requires expensive and difficult procedures. This obstacle can be overcome by obtaining realistic human measurements from 3D human meshes. However, a) there are no well established methods to calculate HBDs from 3D meshes and b) there are no benchmarks to fairly compare results on the HBDE task. Our contribution is twofold. On the one hand, we present a method to calculate right and left arm length, shoulder width, and inseam (crotch height) from 3D meshes with focus on potential medical, virtual try-on and distance tailoring applications. On the other hand, we use four additional body dimensions calculated using recently published methods to assemble a set of eight body dimensions which we use as a supervision signal to our Neural Anthropometer: a convolutional neural network capable of estimating these dimensions. To assess the estimation, we train the Neural Anthropometer with synthetic images of 3D meshes, from which we calculated the HBDs and observed that the network's overall mean estimate error is $20.89$ mm (relative error of 2.84\%). The results we present are fully reproducible and establish a fair baseline for research on the task of HBDE, therefore enabling the community with a valuable method.