3D Mesh Generation from a Defective Point Cloud using Style Transformation
Kenshiro Tamata, Tomohiro Mashita · 2022
In this paper, we propose a method to generate a 3D mesh by inpainting a point cloud having defective points. Our method learns to transfer the style of a defect-free point cloud for a particular object category to a mesh for that object. Then our method applies style transfer from an input point cloud with defective points to a defect-free mesh. Different from previous methods for point cloud inpainting, our method does not directly learn inpainting of defective parts by using set of defective and defect-free point clouds. In our method, we combine an autoencoder and style transformation. We build a 3D mesh generation model from a point cloud by incorporating an inpainting module that inpaints defective parts to the autoencoder. The inpainting module is designed to convert features of defective point clouds to features of defect-free point clouds. Experimental results show that our method can create a 3D model that completes the missing or defective parts of a point cloud.