SAGA-Net: Efficient Pointcloud Completion with Shape-Assisted Graph Attention Neural Network

Abstract

In this paper, we propose a real-time shape-assisted graph attention neural network to perform local pointcloud repairment. The orderless pointclouds require an effective shape encoder to distill local and global geometric feature descriptors. Previous work has attempted to convert pointcloud representation into a voxelized shape or perform grid-transformations. While these approaches can subsequently allow common convolution operations on the structured data, they either pose additional computational cost or disrupt the local geometric information. We present SAGA-Net, an efficient graph attention neural network framework with a prior shape inquiry protocol that effectively extracts local geometric information, locates the descriptor for the missing region and accurately reconstructs the local region in a real-time manner. Our framework has been validated on a benchmark dataset, ShapeNet. We demonstrate that our framework can repair each partial pointcloud with accuracy surpassing other frameworks in most object categories, and improve the computational efficiency by orders of magnitude in terms of time cost.