Unsupervised point cloud completion via pose-aware generation of incomplete point clouds

Abstract

Point clouds are a commonly used data type for representing 3D shapes, but they are often incomplete in the real world. To address this issue, researchers have developed point cloud completion methods for completing missing or incomplete data. These methods have typically been designed for use in either fully-supervised or unpaired settings, but both approaches require the use of complete point clouds, which can be difficult to obtain in practice. In this paper, we propose a novel point cloud completion method that can be trained using only incomplete point clouds and their corresponding poses. Our method extracts a pose-invariant shape feature from the incomplete point cloud, which is shared among all point clouds of the same object. Then, using this shape feature and a sampled pose, we generate a pose-aware incomplete point cloud that appears as if it was captured from that pose. We integrate these incomplete point clouds of various poses into a pseudo-complete point cloud and use it as a ground truth for training a new point cloud completion network. Through extensive experiments, including ablation studies and real point cloud tests, we demonstrate that our method achieves comparable completion performance to existing methods, without requiring any complete point clouds.