Inception Network for Anthropometric Body Measurements Estimation from Structured Point Clouds

Abstract

The need for an automatization in anthropometric body measurements estimation mainly arises from the time-consuming and inefficient process of manual tape measuring of a human body. An automated method for computing the body measurements makes the applications more easily accessible to the end-users, without a need to be familiar with the measuring standards. The lack of standardization is also a big drawback of existing methods in the field. To address the task of body measurements estimation from visual data, we present an end-to-end deep learning-based model built from inception modules to accurately extract values for a pre-defined set of body dimensions from a single grid-structured point cloud. Such data structure jointly combines the benefits of a grid structure as present in 2D images, and the 3D real world coordinates contained in 3D data. We evaluate our method on a large-scale dataset of virtually scanned synthesized human bodies, while reaching the mean absolute test error of approximately 4.5 mm.