Dress Anyone

Abstract

3D garment retargeting for digital characters and avatars involves non-rigid deformation of a 3D garment mesh to plausibly fit the target body mesh in a different pose. Existing neural methods for garment simulation/draping make assumption that the 3D garment is initially fitted over the 3D body, and generally require a canonicalized representation of garments, limiting them to parametric settings. In this paper, we present a novel approach to achieve 3D garment retargeting under non-parametric settings. We propose a novel isomap-based representation to first estimate robust correspondences between garment and body mesh to achieve an initial coarse retargeting, followed by a fast and efficient neural optimization, governed by Physics-based constraints. The proposed framework enables a fast inference pipeline and quick optimization for any 3D garment. We perform extensive experiments on publicly available datasets and our new dataset of 3D clothing and report superior quantitative and qualitative results in comparison to SOTA methods, while demonstrating new capabilities.

Methodology

Architecture

Embedding Calculations

Results

Loose Garments

Human Meshes

3D Bicar Charactors

Layered Garments

Internet Images

Comparison

DIG

Drapenet

Our Dataset

BibTex

@inproceedings{naik2024dressmeupdatasetmethod,
      title={Dress-Me-Up: A Dataset & Method for Self-Supervised 3D Garment Retargeting}, 
      author={Shanthika Naik and Astitva Srivastava and Kunwar Singh and Amit Raj and Varun Jampani and Avinash Sharma},
      year={2024},
      eprint={2401.03108},
      archivePrefix={arXiv},
      primaryClass={cs.CV} 
}