HVPUNet: Hybrid-Voxel Point-Cloud Upsampling Network

Input hybrid voxel (128 ) 3 Output hybrid voxel (128 ) 3 Output hybrid voxel (256 ) 3 Figure 1. We propose an end-to-end technique, HVPUNet, for upsampling 3D point clouds. Unlike most existing techniques, it can upsample points at precise positions with lower computational cost by using hybrid voxels. Our hybrid voxel representation encodes continuous point locations in a structured voxel grid, allowing both precise and efficient reconstruction. HVPUNet consists of two parts: shape completion which imputes missing geometry in the sparse 3D input, and super-resolution which generates a more detailed reconstruction. The estimated hybrid voxel output with normal vectors can be used to reconstruct 3D surfaces.