Normal Estimation for Accurate 3D Mesh Reconstruction with Point Cloud Model Incorporating Spatial Structure.

Taisuke Hashimoto, Masaki Saito

Abstract: In this paper, we propose a network that can accurately infer normal vectors from a point cloud without sacrificing inference speed. The key idea of our model is to introduce a voxel structure to extract spatial features from a given point cloud. Specifically, unlike the other existing methods directly exploiting point clouds, our model leverages two subnetworks called a Opoint networkO and a Ovoxel networkO. The point network extracts local features of a surface from a point cloud, whereas the voxel network transforms the point cloud into voxels and encodes the spatial features from them. The experimental results demonstrate the effectiveness of our method.