Exploring the Properties of Points Generation Network

Abstract

With the development of deep learning, learning-based 3D reconstruction has attracted a substantial amount of attention and various single-image 3D reconstruction networks have been proposed. However, due to self-occlusion, the information captured in a single image is highly limited, resulting in inaccuracy and instability in reconstruction results. In this paper, a feature combination module is proposed to enable existing single-image 3D reconstruction networks to perform 3D reconstruction from multiview images. In addition, we study the impact of the number of the input multiview images as well as the network output points on reconstruction quality, in order to determine the required number of the input multiview images and the output points for reasonable reconstruction. In experiment, point cloud generations with different number of input images and output points are conducted. Experimental results show that the Chamfer distance decreases by 20%∼30% with the optimal number of input multiview images of five and at least 1000 output points.