Deep Patch-based Global Normal Orientation. (September 2022)

• Record Type: Journal Article
• Authors: Wang, Shiyao; Liu, Xiuping; Liu, Jian; Li, Shuhua; Cao, Junjie
• Is Part Of: Computer aided design. Volume 150(2022)
• Publication Date: 2022-09
• Subjects: Point cloud -- Oriented normal estimation -- Deep learning; Computer-aided design -- Periodicals; Engineering design -- Data processing -- Periodicals; Computer graphics -- Periodicals; Conception technique -- Informatique -- Périodiques; Infographie -- Périodiques; Computer graphics; Engineering design -- Data processing; Periodicals; Electronic journals; 620.00420285
• DOI: 10.1016/j.cad.2022.103281
• Languages: English
• ISSNs: 0010-4485

Abstract

Abstract: The accuracy and consistency of point cloud normals are both vital for various successful applications. Great progresses have been achieved recently for accurate unoriented normal estimation by deep learning. However, inferring global consistent orientation directly with deep networks, just as many deep orienters have done, leads to unsatisfactory results since these work only focused on the local geometry of each point. In this paper, we propose a multi-task network which can aggregate global and local information for consistent normal orientation. Specifically, given a query point, a local branch and a global branch take a local patch and a coarse global sub-sampling relative to the query point as input, respectively. Then an unoriented normal and an oriented normal are inferred from the local and global network branches, respectively. Normal estimation from the local branch, as an auxiliary task, boosts the performance of normal orientation. Since the sparse global sample set is as lightweight as a local patch, our network is patch-based. Thus it can be trained on small datasets and be applied to large-scale point clouds straightforwardly. While previous deep orienters whose inputs are whole point clouds cannot achieve it. Extensive experiments on multiple synthetic datasets and raw scanning data demonstrate that our algorithm outperforms the state-of-the-art methods. Our source code, pre-trained model and datasets are available at https://github.com/joycewangsy/DPGO .