The 4D-DRESS dataset aims to advance research in human clothing by providing realistic and challenging real-world data. It contains 520 motion sequences capturing 64 human outfits, amounting to a total of 78k frames. Each frame consists of multi-view images, an 80k-face 3D mesh with vertex-level semantic annotations, and a 1k-resolution texture map.
To create this dataset, the authors developed a semi-automatic 4D human parsing pipeline that efficiently combines automation with human-in-the-loop processes to accurately label the complex 4D scans. This pipeline achieves high-quality vertex-level annotations, with only 1.5% of vertices requiring manual rectification.
The dataset offers diverse garment types, including 4 dresses, 30 upper, 28 lower, and 32 outer garments, captured in dynamic motions. The authors quantify the clothing deformations by computing the mean distances from the garments to the registered SMPL body surfaces, which can reach up to 14.76 cm, highlighting the challenging nature of the dataset.
4D-DRESS serves as a valuable resource for various computer vision and graphics tasks, including clothing simulation, reconstruction, and human parsing. The authors establish several benchmarks to evaluate the performance of state-of-the-art methods on these tasks, revealing the limitations of existing approaches in handling the realistic and complex clothing deformations captured in the dataset.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Wenbo Wang,H... at arxiv.org 04-30-2024
https://arxiv.org/pdf/2404.18630.pdfDeeper Inquiries