GHUM &amp; GHUML: Generative 3D Human Shape and Articulated Pose Models

Xu, Hongyi; Bazavan, Eduard Gabriel; Zanfir, Andrei; Freeman, William T.; Sukthankar, Rahul; Sminchisescu, Cristian

doi:10.1109/cvpr42600.2020.00622

Cited by 301 publications

(146 citation statements)

References 35 publications

Supporting

Mentioning

145

Contrasting

Unclassified

Order By: Relevance

“…Surface meshes are the most commonly used representation for human shape due to their efficiency and compatibility with graphics engines. Not only human body models [2,42,49,75] but also various clothing models leverage 3D mesh representations as separate mesh layers [17,26,27,37,53,67] or displacements from a minimally clothed body [8,45,48,69,74,79]. Recent advances in deep learn-ing have improved the fidelity and expressiveness of meshbased approaches using graph convolutions [45], multilayer perceptrons (MLP) [53], and 2D convolutions [29,37].…”

Section: Related Workmentioning

confidence: 99%

SCALE: Modeling Clothed Humans with a Surface Codec of Articulated Local Elements

Saito

Yang

et al. 2021

2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

View full text Add to dashboard Cite

Section: Related Workmentioning

confidence: 99%

SCALE: Modeling Clothed Humans with a Surface Codec of Articulated Local Elements

Saito

Yang

et al. 2021

2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

View full text Add to dashboard Cite

“…Model-based methods make use of some parametric model as a statistical shape prior. In the literature, some different shape models have been proposed, starting from the SMPL [9], to GHUM [26], which extends the SMPL with face and hand models, or the STAR [27], which is capable of modeling local body shape deformations. Many methods develop on the top of such models.…”

Section: Related Workmentioning

confidence: 99%

Monocular 3D Body Shape Reconstruction under Clothing

2021

View full text Add to dashboard Cite

Estimating the 3D shape of objects from monocular images is a well-established and challenging task in the computer vision field. Further challenges arise when highly deformable objects, such as human faces or bodies, are considered. In this work, we address the problem of estimating the 3D shape of a human body from single images. In particular, we provide a solution to the problem of estimating the shape of the body when the subject is wearing clothes. This is a highly challenging scenario as loose clothes might hide the underlying body shape to a large extent. To this aim, we make use of a parametric 3D body model, the SMPL, whose parameters describe the body pose and shape of the body. Our main intuition is that the shape parameters associated with an individual should not change whether the subject is wearing clothes or not. To improve the shape estimation under clothing, we train a deep convolutional network to regress the shape parameters from a single image of a person. To increase the robustness to clothing, we build our training dataset by associating the shape parameters of a “minimally clothed” person to other samples of the same person wearing looser clothes. Experimental validation shows that our approach can more accurately estimate body shape parameters with respect to state-of-the-art approaches, even in the case of loose clothes.

show abstract

“…Parametric 3D human body models [37,61] are often represented by polygonal meshes and have been widely used to estimate human pose and shape from images and videos [17,28,33], create training data for machine learning algorithms [22,49] and synthesize realistic human bodies in 3D digital environments [68,69]. However, the meshbased representation often requires a fixed topology and lacks flexibility when combined with deep neural networks where back-propagation through the 3D geometry representation is desired.…”

Section: Introductionmentioning

confidence: 99%

“…In this work, we aim to learn articulated neural occupancy representations for various human body shapes and poses. We take inspiration from the traditional mesh-based parametric human body models [37,61], where identityand pose-dependent body deformations are modeled in a canonical space, and then Linear Blend Skinning (LBS) functions are applied to deform the body mesh from the canonical space to a posed space. Analogously, given a set of bone transformations that represent the joint locations and rotations of a human body in a posed space, we first map 3D query points from the posed space to the canonical space via learned inverse linear blend skinning (LBS) functions and then compute the occupancy values via an occupancy network that expresses differentiable 3D body deformations in the canonical space.…”

Section: Introductionmentioning

confidence: 99%