Multi-Scope Feature Extraction for Point Cloud Completion

Abstract

Point cloud completion aims to predict a complete geometric shape based on a partial point cloud. Recent methods often adopt an encoder-decoder framework, where the encoder extracts global features from the partial points and the decoder utilizes a folding-based model to reform multiple 2D grids to 3D surfaces. To effectively explore local features in the partial points, we propose a multi-scope feature extraction method in the encoder, where multiple k-nearest neighbors are considered in the edge convolution. Furthermore, we integrate the original partial point cloud in the decoder to maintain the given geometric shape information. Finally, we refine those coarse points from the decoder by both the merging and sampling operations to output the final completed point cloud. Extensive experiments verify the effectiveness of the proposed approach where both the multi-scope feature extraction and the integration of partial point cloud improve the performance. Overall, our method achieves better performance than the existing methods in both the Earth Mover's Distance (EMD) and the F-score.