We employ contrastive learning to extract complete point cloud structures from partial (incomplete) point clouds for guiding point cloud completion, achieving state-of-the-art (SOTA) results in the field of self-supervised point cloud completion.

A dual-diffusion pretraining framework for point clouds that predicts patch centers and masked geometry to learn robust semantic and geometric representations.

IPFormer introduces context-adaptive instance proposals for vision-based 3D Panoptic Scene Completion, outperforming the state-of-the-art and reducing runtime by over 14x.