【Paper Read】【English】EVA3D: COMPOSITIONAL 3D HUMAN GENERATION FROM 2D IMAGE COLLECTIONS

What can we know from the abstract?

EVA3D, which is an unconditional 3D human generative model learned from 2D image collections only. 

The model can generate 3D humans with detailed geometry and render high-quality images. It is based on a compositional human NeRF representation, which divides the human body into local parts. The model is able to accommodate for the characteristics of sparse 2D human image collections by using a pose-guided sampling strategy. The proposed model achieves state-of-the-art 3D human generation performance regarding both geometry and texture quality. The paper highlights the potential and scalability of EVA3D for "inverse-graphics" diverse human bodies with a clean framework. 

A link is provided to the project page for more information.https://hongfz16.github.io/projects/EVA3D.html

What can we know from the introduction?

The authors propose EVA3D, an unconditional 3D human generative model that utilizes a compositional human NeRF representation to divide the human body into 16 parts, allowing for efficient rendering and high-resolution generation. The authors also propose a pose-guided sampling strategy to address the issues with the imbalanced pose distribution of 2D human image collections. They use SMPL model to guide the transformation between canonical and observation spaces and leverage its human prior during training to better learn 3D human geometry. 

METHODOLOGY

COMPOSITIONAL HUMAN NERF REPRESENTATION

First, select the sampling rays that pass through the human bounding box. Second, twist the sampling points into standard space. Then, study each different part of the human body separately. Finally, integrate the network's output results.

3D HUMAN GAN FRAMEWORK

By incorporating the SMPL template geometry as a strong prior, the network can better learn the correct 3D shape information that is consistent with human anatomy and pose variations.
I don't familar with GAN method yet, so this paragraph will be filled in the future.