3D Garment Digitisation for Virtual Wardrobe Using a Commodity Depth Sensor

Abstract

A practical garment digitisation should be efficient and robust to minimise the cost of processing a large volume of garments manufactured in every season. In addition, the quality of a texture map needs to be high to deliver a better user experience of VR/AR applications using garment models such as digital wardrobe or virtual fitting room. To address this, we propose a novel pipeline for fast, low-cost, and robust 3D garment digitisation with minimal human involvement. The proposed system is simply configured with a commodity RGB-D sensor (e.g. Kinect) and a rotating platform where a mannequin is placed to put on a target garment. Since a conventional reconstruction pipeline such as Kinect Fusion (KF) tends to fail to track the correct camera pose under fast rotation, we modelled the camera motion and fed this as a guidance of the ICP process in KF. The proposed method is also designed to produce a high-quality texture map by stitching the best views from a single rotation, and a modified shape from silhouettes algorithm has been developed to extract a garment model from a mannequin.