Weighted Multi-projection: 3D Point Cloud Denoising with Estimated Tangent Planes

Chaojing Duan, Siheng Chen, Jelena KovacevicarXiv 2018 · OpenReview

TL;DR

This paper introduces "Weighted Multi-Projection" (WMP), a novel algorithm for denoising 3D point clouds. WMP improves upon existing methods by estimating local tangent planes at each point and then averaging projections onto these planes, leading to cleaner point clouds with better preservation of geometric details.

What Problem Does This Paper Address?

3D point clouds, essential for applications like robotics and augmented reality, are often plagued by noise due to sensor limitations. This noise degrades the accuracy of subsequent tasks like surface reconstruction and object recognition. Existing denoising methods struggle with issues like over-smoothing, clustering, and deformation. This paper aims to address these challenges by developing a denoising technique that effectively removes noise while preserving the underlying geometric structure of the point cloud.

Methods

The proposed Weighted Multi-Projection (WMP) algorithm operates in three main stages:

  1. Graph Construction: A graph is built where nodes represent 3D points, and edges connect neighboring points based on their Euclidean distance. This graph captures the local and global geometric structure of the point cloud.
  2. Tangent Plane Estimation: For each point, a local tangent plane is estimated. This is achieved by minimizing the weighted sum of squared distances from the point and its neighbors to the plane, using the graph structure. The normal vector of the tangent plane is derived from the eigenvector corresponding to the smallest eigenvalue of a weighted covariance matrix.
  3. Weighted Multi-Projection: Instead of a single projection, each point is projected onto its own estimated tangent plane and the tangent planes of its neighbors. The final denoised point is obtained by a weighted average of these multiple projections. This "multi-projection" strategy is designed to be more robust to noise than a single projection.

Datasets

The authors evaluated their algorithm on the ShapeNetCore dataset, which contains 55 manually verified common object categories with over 50,000 clean 3D mesh models. They sampled points from these models, added Gaussian noise to create noisy point clouds, and then compared their method against others.

Key Insights and Ideas

  • Tangent Planes as Geometric Priors: The core idea is that local tangent planes can effectively approximate the underlying manifold of a point cloud, and projecting noisy points onto these planes can remove noise components orthogonal to the surface.
  • Multi-Projection for Robustness: Averaging projections from multiple tangent planes (the point's own and its neighbors') is shown to be more robust and less prone to deformation than a single projection onto one plane. This is because neighboring points often share similar geometric information, and their tangent planes are inherently similar.
  • Theoretical Guarantees: The paper provides theoretical analysis for surface normal estimation, achieving a tighter bound than previous work. They also theoretically demonstrate that multi-projection is superior to one-time projection.

Key Results

  • Empirical Performance: WMP consistently outperformed state-of-the-art denoising algorithms (Bilateral Filter, Graph-Based Denoising, PDE-based filter, Non-local Denoising) across all nine tested categories in the ShapeNetCore dataset.
  • Evaluation Metrics: The algorithm showed superior performance in terms of Signal-to-Noise Ratio (SNR), Mean Squared Error (MSE), and Chamfer Distance (CD). Figures 4, 5, and 6 visually confirm these improvements.
  • Qualitative Results: Figure 1 demonstrates that WMP produces cleaner and smoother mesh reconstructions compared to other methods, preserving finer details.
  • Computational Complexity: The algorithm has a computational complexity of O(N(log N + k)), where N is the number of points and k is the average number of neighbors, making it efficient for practical use.

Limitations and Future Work

  • Theoretical Generalization: The theoretical analysis for surface normal estimation and projection superiority was primarily shown for a 2D case (semicircle). Generalizing these results to 3D point clouds (e.g., spheres) is left for future work.
  • Parameter Tuning: While the paper mentions tuning parameters for comparison algorithms, the optimal parameter selection for WMP itself (e.g., the $\epsilon$ for $\epsilon$-NN graph) is not explicitly detailed for all scenarios.
  • Sharp Edges: The paper notes that performance slightly degrades on datasets with more sharp edges (like chairs and sofas), suggesting that preserving sharp features remains a challenge for many denoising algorithms.

Summary updated

Abstract

As a collection of 3D points sampled from surfaces of objects, a 3D point cloud is widely used in robotics, autonomous driving and augmented reality. Due to the physical limitations of 3D sensing devices, 3D point clouds are usually noisy, which influences subsequent computations, such as surface reconstruction, recognition and many others. To denoise a 3D point cloud, we present a novel algorithm, called weighted multi-projection. Compared to many previous works on denoising, instead of directly smoothing the coordinates of 3D points, we use a two-fold smoothing: We first estimate a local tangent plane at each 3D point and then reconstruct each 3D point by weighted averaging of its projections on multiple tangent planes. We also provide the theoretical analysis for the surface normal estimation and achieve a tighter bound than in a previous work. We validate the empirical performance on the dataset of ShapeNetCore and show that weighted multi-projection outperforms its competitors in all nine classes.
BibTeX
@misc{duan2018weighted,
  author = {Chaojing Duan and Siheng Chen and Jelena Kovacevic},
  title = {{Weighted Multi-projection: 3D Point Cloud Denoising with Estimated Tangent Planes}},
  year = {2018},
  eprint = {1807.00253},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1807.00253},
}
Figure