[Paper Review] PointCleanNet: Learning to Denoise and Remove Outliers from Dense Point Clouds

Marie‐Julie Rakotosaona, Vittorio La Barbera+3 more|UCL Discovery (University College London)|Jan 4, 2019

Optical measurement and interference techniquesComputer Science50 references36 citations

TL;DR

PointCleanNet uses a two-stage deep learning approach to first remove outliers and then denoise dense point clouds, improving robustness across varying noise levels without user-tuned parameters.

ABSTRACT

Point clouds obtained with 3D scanners or by image-based reconstruction techniques are often corrupted with significant amount of noise and outliers. Traditional methods for point cloud denoising largely rely on local surface fitting (e.g., jets or MLS surfaces), local or non-local averaging, or on statistical assumptions about the underlying noise model. In contrast, we develop a simple data-driven method for removing outliers and reducing noise in unordered point clouds. We base our approach on a deep learning architecture adapted from PCPNet, which was recently proposed for estimating local 3D shape properties in point clouds. Our method first classifies and discards outlier samples, and then estimates correction vectors that project noisy points onto the original clean surfaces. The approach is efficient and robust to varying amounts of noise and outliers, while being able to handle large densely-sampled point clouds. In our extensive evaluation, both on synthesic and real data, we show an increased robustness to strong noise levels compared to various state-of-the-art methods, enabling accurate surface reconstruction from extremely noisy real data obtained by range scans. Finally, the simplicity and universality of our approach makes it very easy to integrate in any existing geometry processing pipeline.

Motivation & Objective

• Motivation to reduce reliance on parameter-tuned traditional denoising methods for diverse noise in 3D point clouds.
• Develop a data-driven, robust cleaning pipeline that handles both outliers and noise in unordered, dense point clouds.
• Leverage a two-stage network to separate outlier removal from denoising and enable iterative refinement.
• Demonstrate improved denoising and outlier removal compared to state-of-the-art baselines on synthetic and real data.

Proposed method

• Two-stage architecture: first a local outlier detection network to discard outliers, then a second network to estimate per-point displacement vectors for denoising.
• Based on an adaptation of PCPNet architecture to operate on local patches around each point.
• Loss functions balance proximity to the surface with regularization to maintain even point distribution and avoid clustering.
• Iteration of the denoising network is used to progressively reduce residual noise and improve convergence.
• Training uses synthetically corrupted paired noisy-clean data to learn to classify outliers and predict displacements.
• The approach is permutation- and transform-invariant and designed to work on dense point clouds without heavy parameter tuning.

Experimental results

Research questions

• RQ1 Can a two-stage, data-driven network reliably remove outliers and denoise dense point clouds without user-specified noise models?
• RQ2 Does iterative application of PointCleanNet improve denoising performance across varying noise levels and artefacts?
• RQ3 How does PointCleanNet compare to traditional and learning-based baselines in denoising accuracy and outlier detection across synthetic and real datasets?
• RQ4 Is the method robust to different surface types and capable of preserving high-curvature features without extra surface information?

Key findings

• PointCleanNet shows improved denoising performance over a range of noise levels, outperforming several baselines when iterated.
• The method demonstrates robustness to strong noise and outliers, with better stability across noise variations compared to state-of-the-art methods.
• Outlier removal and denoising are achieved without user-tuned parameters, leveraging learned local patches and an adapted loss.
• Qualitative and quantitative results indicate lower residual error on real-world scanning data compared to competing methods.
• The approach integrates easily into existing geometry processing pipelines and is suitable for dense point clouds.