3D point cloud is an important 3D representation for capturing real world 3D. However, real-scanned 3D point clouds are often incomplete, and it is to recover complete point clouds for downstream applications. Most point cloud completion methods use Chamfer Distance (CD) loss for. The CD loss estimates correspondences between two point clouds by nearest neighbors, which does not capture the overall point density on the generated shape, and therefore likely leads to non-uniform cloud generation. To tackle this problem, we propose a novel Point-Refinement (PDR) paradigm for point cloud completion. PDR consists of Conditional Generation Network (CGNet) and a ReFinement Network (RFNet). The uses a conditional generative model called the denoising diffusion model (DDPM) to generate a coarse completion conditioned on the observation. DDPM establishes a one-to-one pointwise mapping between generated point cloud and the uniform ground truth, and then optimizes the squared error loss to realize uniform generation. The RFNet refines the output of the CGNet and further improves quality of the completed point. Furthermore, we develop a novel dual-path architecture for both. The architecture can (1) effectively and efficiently extract-level features from partially observed point clouds to guide completion, (2) accurately manipulate spatial locations of 3D points to obtain smooth and sharp details. Extensive experimental results on various benchmark show that our PDR paradigm outperforms previous state-of-the-art for point cloud completion. Remarkably, with the help of the RFNet, we accelerate the iterative generation process of the DDPM by up to 50 times much performance drop.
No takes yet. Share an insight, caveat, or question.
Lyu et al. (2021) studied this question.