3D Point Cloud Completion Based on Multi-Scale Degradation

Abstract

Recent advances in 3D point cloud completion adopt unsupervised deep learning-based methods, which does not rely on labeled data and improves generalization ability. However, existing methods tend to focus more on the generation overall shape rather than detailed structure. To explore unsupervised 3D point cloud completion methods that give attention to both, we propose a multi resolution completion net (MRC-Net) which introduces a multi-scale degradation (KM- mask) and multi-discriminator into GAN inversion paradigm. First, degrade the point clouds completed by the generator under three different resolutions. Then, the losses of multi-stage reconstruction and feature matching by multi-scale discriminator are used to jointly optimize the generator. The experimental results demonstrate that MRC-Net outperforms existing unsupervised point cloud completion methods and has better completion performances on virtual scanning dataset.