ConDor: Self-Supervised Canonicalization of 3D Pose for Partial Shapes

Abstract

is a self-supervised method that learns to Canonicalize the 3D orientation and position (3D pose) for full and partial shapes. (left) Our method takes un-canonicalized 3D point clouds (gray) from different categories as input and produces consistently canonicalized outputs (colored). (right) Our method can also operate on partial point clouds (missing part of shape shown only for visualization). In addition, ConDor can also learn consistent co-segmentation of shapes without supervision, visualized as colored parts.