Learning Contours for Point Cloud Completion

The integrity of a point cloud frequently suffers from discontinuous material surfaces or coarse sensor resolutions. Existing methods focus on reconstructing the overall structure, but salient points or small irregular surfaces are difficult to be predicted. Toward this issue, we propose a new end-to-end neural network for point cloud completion. To avoid non-uniform point density, the regular voxel centers are selected as reference points. The encoder and decoder are designed with Patchify, transformers, and multilayer perceptrons. An implicit classifier is incorporated in the decoder to mark the valid voxels that are allowed for diffusion after removing vacant grids from completion. With newly designed loss function, the classifier is trained to learn the contours, which helps to identify the grids that are difficult to be judged for diffusion. The effectiveness of the proposed model is validated in the experiments on the indoor ShapeNet dataset, the outdoor KITTI dataset, and the airbone laser dataset by competing with state-of-the-art methods, which show that our method can predict more accurate point coordinates with rich details and uniform point distributions.

Automated summary

This paper presents a new end-to-end neural network for point cloud completion. The proposed method selects regular voxel centers as reference points and designs the encoder and decoder with Patchify, transformers, and multilayer perceptrons. An implicit classifier is incorporated in the decoder to mark valid voxels for diffusion. The effectiveness of the proposed model is validated through experiments on various datasets, showing improved accuracy and detail in predicting point coordinates with uniform distributions.