Regional dynamic point cloud completion network

Abstract

Point cloud completion network often encodes points into a global feature vector, then predicts the complete point cloud through the vector generation process. However, this method may not accurately capture complex shapes, as global feature vectors struggle to recover their detailed structure. In this paper, we present a novel shape completion network, namely RD-Net, that innovatively focuses on the interaction of information between points to provide both local and global information for generating fine-grained complete shape. Specifically, we propose a stored iteration-based method for point cloud sampling that quickly captures representative points within the point cloud. Subsequently, in order to better predict the shape and structure of the missing part, we design an iterative edge-convolution module. It uses a CNN-like hierarchy for feature extraction and learning context information. Moreover, we design a two-stage reconstruction process for latent vector decoding. We first employ a feature-points-based multi-scale generating decoder to estimate the missing point cloud hierarchically. This is followed by a self-attention mechanism that refines the generated shape and effectively generates structural details. By combining these innovations, RD-Net achieves a 2% reduction in CD error compared to the state-of-the-art method on the ShapeNet-part dataset.