A survey of 3D object reconstruction methods

Abstract

In this work, we provide a state-of-the-art survey of deep learning-based single- and multi-view 3D object reconstruction methods. In a broad sense, 3D reconstruction methods take single or multiple 2D images to model shapes with different representations such as: voxels, meshes, point clouds and implicit functions. In this paper, the methods are grouped based on their shape representations and are presented in detail with their deep neural network architectures, supervision mechanisms and reconstruction accuracies on benchmark datasets.