Date: April 2024.
Source: arXiv:2404.02686v1 [cs.CV]. https://doi.org/10.48550/arXiv.2404.02686.
Research Summary: Digital avatar research has increasingly shifted toward modeling, animating, and reconstructing clothed human representations as a key step toward realistic avatars. Existing 3D cloth generation methods are typically garment-specific or trained entirely on synthetic data, limiting fine detail and realism. Design2Cloth addresses this with a high-fidelity 3D generative model trained on DigitalMe, a real-world dataset of more than 2,000 subject scans. The team built a user-friendly adversarial model capable of generating diverse, detailed clothing from a simple 2D cloth mask drawing. Qualitative and quantitative experiments show Design2Cloth outperforms current state-of-the-art cloth generative models by a large margin, and the method also achieves high-quality clothing reconstructions from single in-the-wild images and 3D scans.
3dMD’s Role: Design2Cloth’s training data, named DigitalMe, comprises 2,010 real-world clothed human scans captured on a 3dMDbody System, 14 cameras and 12 uniform LED lighting panels, at approximately 150,000 vertices per scan. Subjects spanned a wide range of ages, body types, and ethnicities, wearing a variety of garments across multiple poses starting from a canonical pose. From these raw scans, an automated pipeline fits the SMPL body model and extracts individual clothing items using multi-view rendering and segmentation, producing 2,010 unique cloth meshes used to train the model.
Article: Design2Cloth: 3D Cloth Generation from 2D Masks.
Authors: Jiali Zheng, Rolandos Alexandros Potamias, Stefanos Zafeiriou. Imperial College London.