GeoHi-GNN: Geometry-aware hierarchical graph representation learning for normal estimation

• Published in: Computer aided geometric design 2024-11, Vol.114, p.102390, Article 102390
• Main Authors: Li, Nannan, Li, Xinyuan, Zhou, Jun, Jiang, Dong, Liu, Jian, Qin, Hong
• Format: Article
• Language: English
• Subjects: 3D deep learning; Dynamic graph convolution; Hierarchical graphs; Normal estimation; Representation learning; Surface fitting
• ISSN: 0167-8396
• DOI: 10.1016/j.cagd.2024.102390

Normal estimation has been one of the key tasks in point cloud analysis, while it is challenging when facing with severe noises or complex regions. The challenges mainly come from the selection of supporting points for estimation, that is, improper selections of points and points' scale will lead to insufficient information, loss of details, etc. To this end, this paper proposes one feature-centric fitting scheme, GeoHi-GNN, by learning geometry-aware hierarchical graph representation for fitting weights estimation. The main functional module is the continuously conducted Hierarchically Geometric-aware (HG) module, consisting of two core operations, namely, the graph node construction (GNC) and the geometric-aware dynamic graph convolution (GDGC). GNC aims to aggregate the feature information onto a smaller number of nodes, providing global-to-local information while avoiding the interferences from noises in larger scales. With these nodes distributed in different scales, GDGC dynamically updates the node features regarding to both intrinsic feature and extrinsic geometric information. Finally, the hierarchical graphical features are cascaded to estimate the weights for supporting points in the surface fitting. Through the extensive experiments and comprehensive comparisons with the state-of-the-arts, our scheme has exhibited many attractive advantages such as being geometry-aware and robust, empowering further applications like more accurate surface reconstruction. •Proposing a hierarchically geometric-aware fitting scheme for normal estimation on point cloud.•Proposing HG module for aggregating and updating features when shrinking graph scale.•Introducing patch-perceptual loss to infer the noise level and reweight the supporting points.