Yang Juyoung ; Ahn Pyunghwan ; Kim Doyeon ; Lee Haeil ; Kim Junmo

Abstract

With the development of 3D scanning technologies, 3D vision tasks have becomea popular research area. Owing to the large amount of data acquired by sensors,unsupervised learning is essential for understanding and utilizing point cloudswithout an expensive annotation process. In this paper, we propose a novelframework and an effective auto-encoder architecture named "PSG-Net" forreconstruction-based learning of point clouds. Unlike existing studies thatused fixed or random 2D points, our framework generates input-dependentpoint-wise features for the latent point set. PSG-Net uses the encoded input toproduce point-wise features through the seed generation module and extractsricher features in multiple stages with gradually increasing resolution byapplying the seed feature propagation module progressively. We prove theeffectiveness of PSG-Net experimentally; PSG-Net shows state-of-the-artperformances in point cloud reconstruction and unsupervised classification, andachieves comparable performance to counterpart methods in supervisedcompletion.