ShapeBoost proposes a new approach to human shape recovery, utilizing part-based parameterization and clothing-preserving augmentation. The method surpasses state-of-the-art techniques in accurately estimating body shapes under various conditions.
Accurate human shape estimation from monocular RGB images is challenging due to diverse body shapes and clothing variations. ShapeBoost introduces a novel framework that decomposes human shape into bone lengths and part widths, achieving pixel-level alignment even for rare body shapes and high accuracy for people wearing different types of clothes. By adopting a new parameterization technique, the method balances flexibility and validity using a semi-analytical shape reconstruction algorithm. The proposed clothing-preserving data augmentation module generates realistic images with diverse body shapes and accurate annotations. Experimental results demonstrate the superiority of ShapeBoost over existing methods in handling diverse body shapes and varied clothing situations.
Previous approaches have struggled with overfitting on body shape estimation due to limited datasets featuring diverse body shapes. ShapeBoost addresses this limitation by proposing a new parameterization that decomposes human shape into bone lengths and part widths, enabling accurate reconstruction of extreme body shapes while encouraging pixel-level alignment. The method outperforms state-of-the-art techniques in both thick clothes situations and extreme body shape scenarios.
The main contributions of ShapeBoost include an accurate and robust human shape parameterization, a clothing-preserving data augmentation module, and a shape reconstruction framework that outperforms previous approaches in handling diverse clothing as well as extreme body shapes.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Siyuan Bian,... at arxiv.org 03-05-2024
https://arxiv.org/pdf/2403.01345.pdfDeeper Inquiries