TRFit: learning 3D point cloud normal estimation with transformer

Abstract

In this study, we provide an approach named TRFit for unstructured 3D point cloud normal estimation. It handles noise and uneven densities point clouds well. Recently, learning-based normal estimation methods have significantly outperformed traditional methods on benchmark normal estimation datasets. In order to estimate normals, they frequently employed neural networks to learn point-wise weights for weighted least squares polynomial surfaces fitting. However, existing methods often ignore local geometric relationships, which will make the fitted surface significantly different from the real. To this end, we propose to use graph convolutional to learn local structural information. Meanwhile, we suggest the Geometric Relation Transformer (GRT), a transformer-based scale aggregation module, to fully utilize points from various neighborhood sizes. It can adaptively capture the relations between different regions. We achieve state-of-the-art results on the baseline normal estimation dataset, and experimental results show that TRFit obviously improves the accuracy of normal estimates, preserves their details. Moreover, it exhibits robustness to noise, density variations, and outliers. Besides, we demonstrate its application to surface reconstruction and denoising.