A feature-preserving framework for point cloud denoising
Graphical abstract
Introduction
Point cloud data has attracted considerable attention due to the rapid development of scanning devices (e.g., Microsoft Kinect, Xtion Pro, Google Project Tango, and Intel RealSense). However, even with high-fidelity devices, the corruption of scanned data is usually inevitable from the degradation during its acquisition. Thus, recovering high quality point clouds from the noisy inputs is a typical inverse problem in geometry processing. The main challenge is to distinguish sharp features and noise as they are of high frequency information, and at the same time prevent unnatural effects during the denoising process.
Over the years, state-of-the-art methods have been proposed for recovering noise-free point clouds. In [1], Öztireli et al. proposed a method combining moving least square and local kernel regression to preserve geometric features. However, when the noise level increases, it is unlikely to perform well with satisfactory results. Especially, this limitation is more severe for point clouds containing sharp features. By using robust principal component analysis, [2] exploits sparse characteristics of subspaces to preserve sharp features, which is also robust against outliers. Unfortunately, [2] usually over-sharpens curved features because it depends on sparse and low-rank modeling. In [3], Huang et al. developed an edge-aware upsampling technique to recover the point cloud with well-preserved sharp features. Nevertheless, it sometimes flattens fine details, since it chooses large neighborhood size to pull points away from sharp features for upsampling. Moreover, this method may be sensitive to high density noise, because of its unfaithful upsampling around sharp features without explicitly estimating the point normal field. Sun et al. [4] extended minimization to point clouds for preserving sharp features. Although their method achieves impressive results for data with piecewise constant priors, it tends to flatten smoothly curved regions and produce pseudo edges in these regions due to its high requirement of sparsity.
As we have seen, the aforementioned state-of-the-art methods are either less able to preserve sharp features well, or may produce unnatural artifacts in denoised results. To overcome these limitations, we propose a novel feature-preserving framework to recover point normals as well as positions. It consists of three cascaded stages: normal filtering; feature detection; and multi-normal point updating, as illustrated in Fig. 1. In the first stage, we present an anisotropic second order variational method to restore the normal field from the noisy input; see Fig. 1, Fig. 1. Based on the filtered normals, we define normal and point voting tensors, and then introduce a bi-tensor voting scheme to detect features; see Fig. 1(c). The scheme utilizes the best properties of the normal and point tensor voting and overcomes the weakness of both. In the following multi-normal point updating stage, we estimate the multiple normals per feature point using a RANSAC-based algorithm; see Fig. 1(d). Then, we update point positions according to the restored normals (the normals at non-feature points are produced by our second order normal filtering, and those at feature points are estimated by the RANSAC-based algorithm); see Fig. 1(e). The main contributions of the paper are summarized as follows:
- •
An anisotropic second order regularization method is presented to restore the point normal field. It is able to preserve sharp features well and simultaneously prevent unnatural effects in denoised results.
- •
A bi-tensor voting scheme, which combines the normal and point tensor voting, is proposed to detect features on the noisy input. The combined technique is not only robust against noise but also can accurately locate features.
- •
A simple yet effective RANSAC-based algorithm is introduced to estimate the multiple normals at each feature point. It can significantly reduce cross artifacts during the point reconstruction process.
Section snippets
Related work
Point cloud denoising is a fundamental problem in digital geometry processing, which has been studied extensively. It is beyond our scope to review numerous existing methods, and we only review noticeable ones closely related to this work. Generally, we classify point cloud denoising methods into four main categories as following.
MLS-based methods. Moving least squares (MLS) has been originally designed for surface reconstruction by Alexa et al. [5]. Later, many extensions and modifications [6]
Normal filtering via second order regularization
The noisy point cloud can be considered as a set of unorganized points sampled from a 2-dimensional manifold in , where is the number of sampled points. For the th point of , its index set of K-nearest neighborhood (KNN) is denoted as , which consists of elements. We further arrange KNNs of the point in a counterclockwise order using local PCA.
Feature detection by bi-tensor voting
The tensor voting is a fundamental tool in geometry processing for accurately detecting features on high-quality meshes [25], [26]. Recently, it was extended to point clouds by analyzing the normal voting tensor [27], [28] or the point voting tensor [29], [30]. However, performing direct the normal or point tensor voting on noisy point clouds usually produces spurious effects. Specifically, the point tensor voting is sensitive to the noise, which tends to produce pseudo features in smooth
Multi-normal point updating
After obtaining filtered point normals, it is necessary to reposition points to match the filtered normals. Because the single normals at feature points are ambiguous and undefinable, it is necessary to adopt multi-normal point updating strategy as proposed in previous works [30], [31], [32]. This technique can overcome cross artifacts at sharp features and preserve sharp features better than the conventional single normal based approaches, especially when the point cloud is corrupted with high
Experimental results and comparisons
To testify the performance of our point cloud denoising method, we perform it on both synthetic and real data. The tested point clouds are contaminated by either synthetic or raw scanned noise. The synthetic noise is generated by a zero-mean Gaussian function with standard deviation proportional to the diagonal of the axis-aligned bounding box of the ground truth. We also provide visual and quantitative comparisons of our method to the state-of-the-art ones including RIMLS [1], MRPCA [2], EAR
Conclusion
In this paper, a feature-preserving framework has been proposed to recover a noisy-free point cloud. It first utilizes a second order regularization to restore the normal field. With the filtered normals, a well-designed bi-tensor voting scheme is introduced to detect features, which overcomes the weakness of the normal and point tensor voting. Finally, point positions are reconstructed by the multi-normal strategy to reduce cross artifacts. Experimental results show that our denoising method
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
The authors would like to thank anonymous reviewers for their constructive suggestions for improving the manuscript. Zheng Liu’s research is supported by National Natural Science Foundation of China (No. 61702467). Weina Wang’s research is supported by Zhejiang Provincial Natural Science Foundation of China (No. LQ20A010007). Ling Zhang’s research is supported by China Postdoctoral Science Found (No. 2018M642933) and National Natural Science Foundation of China (No. 61902286). Zhong Xie’s
References (32)
- et al.
Denoising point sets via minimization
Comput Aided Geom Design
(2015) - et al.
A novel anisotropic second order regularization for mesh denoising
Comput Aided Geom Des
(2019) - et al.
Feature detection of triangular meshes based on tensor voting theory
Comput Aided Des
(2009) - et al.
Multi-scale tensor voting for feature extraction from unstructured point clouds
Graph Models
(2012) - et al.
Constraint-based point set denoising using normal voting tensor and restricted quadratic error metrics
Comput Graph
(2018) - et al.
Point cloud normal estimation via low-rank subspace clustering
Comput Graph
(2013) - et al.
Rolling normal filtering for point clouds
Comput Aided Geom Design
(2018) - et al.
Feature preserving point set surfaces based on non-linear kernel regression
Comput Graph Forum
(2009) - et al.
Point cloud denoising via moving RPCA
Comput Graph Forum
(2017) - et al.
Edge-aware point set resampling
ACM Trans Graph
(2013)
Computing and rendering point set surfaces
IEEE Trans Vis Comput Graphics
Robust moving least-squares fitting with sharp features
ACM Trans Graph
Point-sampled cell complexes
ACM Trans Graph
Algebraic point set surfaces
ACM Trans Graph
Dynamic sampling and rendering of algebraic point set surfaces
Comput Graph Forum
Parameterization-free projection for geometry reconstruction
ACM Trans Graph
Cited by (47)
Accurate and complete line segment extraction for large-scale point clouds
2024, International Journal of Applied Earth Observation and GeoinformationRobust multi-task learning network for complex LiDAR point cloud data preprocessing
2024, Expert Systems with ApplicationsReconstruction and Preservation of Feature Curves in 3D Point Cloud Processing
2024, CAD Computer Aided DesignAlternately denoising and reconstructing unoriented point sets
2023, Computers and Graphics (Pergamon)Robust and Accurate Feature Detection on Point Clouds
2023, CAD Computer Aided DesignFeature-preserving Mumford–Shah mesh processing via nonsmooth nonconvex regularization
2022, Computers and Graphics (Pergamon)