TreeNet: Structure preserving multi-class 3D point cloud completion

Long Xi, Wen Tang, Tao Ruan Wan

Abstract: Highlights•A novel TreeNet-multiclass decoder is proposed for multi-class 3D point cloud completion.•A novel TreeNet-binary decoder is proposed, which focuses on generating points in missing areas and fully preserving the original partial input 3D point cloud.•A novel TreeNet is proposed, which combines the advantages of the TreeNet-multiclass and the TreeNet-binary for 3D point cloud completion.•Three novel forward and backward propagation methods are proposed to train TreeNet-multiclass, TreeNet-binary and TreeNet decoders, respectively.•TreeNet-multiclass, TreeNet-binary and TreeNet exhibit strong generalization to unknown classes that are never trained.