Deep Learning for Robust Normal Estimation in Unstructured Point Clouds

Abstract

Normal estimation in point clouds is a crucial first step for numerous algorithms, from surface reconstruction and scene understanding to rendering. A recurrent issue when estimating normals is to make appropriate decisions close to sharp features, not to smooth edges, or when the sampling density is not uniform, to prevent bias. Rather than resorting to manually‐designed geometric priors, we propose to learn how to make these decisions, using ground‐truth data made from synthetic scenes. For this, we project a discretized Hough space representing normal directions onto a structure amenable to deep learning. The resulting normal estimation method outperforms most of the time the state of the art regarding robustness to outliers, to noise and to point density variation, in the presence of sharp edges, while remaining fast, scaling up to millions of points.

PCPNet Learning Local Shape Properties from Raw Point Clouds

“…One source of error of our current network compared, e.g. to the results of Boulch et al [BM16] is that our method does not perform as well in the case of changes in sampling density. This is because our network was trained only on uniformly sampled point sets and therefore is not as robust to such changes.…”

“…Comparison of the RMS normal angle error of our method (ss: single scale, ms: mult‐scale, +c: joint normals and curvature) to geometric methods (jet fitting and PCA) with three patch sizes and two deep learning methods (Boulch et al [BM16] and PointNet [QSMG17]). Note that geometric methods require correct parameter settings, such as the patch size, to achieve good results.…”

“…However, such discretizations introduce additional quantization errors leading to poor‐quality local estimates. In an interesting recent work on normal estimation, Boulch et al [BM16] generate a set of local approximating planes, and then propose a data‐driven denoising approach in the resultant Hough‐space to estimate normals. In Section 5, we demonstrate that even such a specialized approach leads to errors on raw point clouds due to additional ambiguities introduced by the choice of the representation (RANSAC‐like Hough space voting).…”

Deep learning for urban remote sensing

“…We propose a novel method [5] for normal estimation in unorganized point clouds, based on a deep neural network. It is robust to noise, outliers, density variation and sharp edges, and it scales well to millions of points.…”

“…The right image details a case of very high density variation, as roofs are much more densely sampled than facade walls. Details on the method and more quantitative results are presented in [5]. …”