Point Cloud Semantic Scene Completion from RGB-D Images

In this paper, we devise a novel semantic completion network, called point cloud scene network (PCSSC-Net), for indoor scenes solely based on clouds. Existing networks still suffer from their inability of fully recovering complex structures and contents global geometric descriptions neglecting hints. To extract infer comprehensive information partial input, design patch-based contextual encoder to hierarchically learn point-level, patch-level, scene-level with divide-and-conquer strategy. Consider that the semantics afford high-level clue constituting geometry an environment, articulate semantics-guided decoder where could help cluster isolated points in latent space complicated geometry. Given fact real-world scans tend be incomplete as ground truth, choose synthesize dataset RGB-D images annotate complete clouds truth supervised training purpose. Extensive experiments validate our new method achieves state-of-the-art performance, contrast current methods applied dataset.