Core Concepts
The authors propose a method that leverages shape and deformation priors trained on synthetic data to accurately capture garment shapes and deformations, enabling the recovery of realistic 3D garment meshes from in-the-wild images.
Abstract
The content discusses a novel approach for recovering realistic 3D garment meshes from images using shape and deformation priors. The method overcomes challenges in modeling loose-fitting clothing by introducing a fitting process that utilizes learned priors to accurately capture garment shapes and deformations. By optimizing pre-trained deformation models and refining mesh vertex positions, the approach outperforms existing methods in terms of reconstruction accuracy. The study includes detailed experiments, implementation details, comparisons with state-of-the-art methods, ablation studies, and more results showcasing the effectiveness of the proposed approach.
Stats
Figure 1. Fitting method leverages shape and deformation priors trained on synthetic data.
ISP model represents garments using 2D panels and 3D surfaces.
Training ISP requires rest state patterns generated by flattening algorithms.
Deformation model predicts occupancy values and corrective displacements.
Two-stage fitting process optimizes parameters of pre-trained deformation model.
Results show significant improvement in reconstruction accuracy compared to baselines.
Quotes
"Our approach can faithfully recover garment mesh from input images."
"Our method outperforms existing methods in terms of Chamfer Distance (CD) and Intersection over Union (IoU)."
"Directly optimizing vertex positions without optimizing the deformation model is not as effective."