Method for estimating real-scale 3D human body shape from an image based on 3D camera calibration and computer graphics-based reverse projection photogrammetry.

Abstract

The combination of computer vision (CV) and computer graphics (CG) is being developed for use in many fields. Consequently, reverse projection photogrammetry, which identifies geometric properties of a subject based on accurate reproduction of the image content, is beginning to replace analysis combining CV and CG. Since an image captured by a camera has two-dimensional (2D) geometry, estimating real-scale three dimensional (3D) information about a human or object from a low-resolution security camera image is a challenge and has not been achieved without prior knowledge of the person or object. However, deep learning technology that applies fitting a 3D human body shape model to a human image has been developed, but it is difficult to scale the reconstructed model to the actual scale with only a 2D image as an input. In this study, we propose a novel method to estimate a real-scale 3D human body shape model (SMPL-X model) from a human image via a combination of 3D camera calibration and CG-based reverse projection photogrammetry. The method estimates the position, orientation, posture, and body shape of a 3D human body shape model of a human image in a non-straight posture, which is difficult to analyze conventionally. The method was also used to estimate height and weight based on the estimated 3D human body shape, greatly expanding the range of analysis of height and weight estimation. The equal error rate from a few hundred to a few thousand comparisons was evaluated toward realizing person verification.