Semantic Point Completion Network for 3D Semantic Scene Completion

Semantic scene completion (SSC) is composed of scenecompletion (SC) and semantic segmentation. Most of the existingmethods carry out SSC in a regular 3D grid space, where 3D CNNscause unnecessary computational cost on empty voxels. In this work,a Semantic Point Completion Network (SPCNet) is proposed toaddress SSC in the point cloud space. Specifically, SPCNet is anEncoder-decoder architecture, in which an Observed Point Encoderis applied to extract the features of observed points, and an Observedto Occluded Point Decoder is responsible for mapping the featuresto the occluded points. Based on the SPCNet, we further introducean Image-point Fused Semantic Point Completion Network (IPFSPCNet), which aims to boost the performance of SSC by combiningthe texture with geometry information. Evaluations are conducted ontwo public datasets. Experimental results show that our method canaddress the SC problem in the point cloud space. Compared to stateof-the-art approaches, our method can achieve satisfying results onthe SSC task.