Deep learning-based solution for 3D reconstruction from single RGB images

Abstract

In this work, we introduce a deep learning framework designed to generate a 3D triangular mesh from a single image. Unlike existing methods that commonly represent 3D shapes using volumes or point clouds, our approach takes a different path. We employ a distinct strategy by utilizing a graph-based convolutional neural network within our network architecture to represent the 3D mesh. This network progressively deforms an ellipsoid based on features extracted from the input image, leveraging perceptual information. To ensure stability in the deformation process, we implement a coarse-to-fine strategy. Through extensive experimentation, we demonstrate the efficacy of our method. It not only producesa qualitatively superior mesh model with enhanced details but also outperforms state-of-the-art approaches in terms of 3D shape estimation accuracy.