[Paper Review] 3D Point Cloud Denoising via Deep Neural Network based Local Surface Estimation

Chaojing Duan, Siheng Chen+1 more|arXiv (Cornell University)|Apr 9, 2019

3D Shape Modeling and Analysis22 references4 citations

TL;DR

This paper proposes Neural Projection Denoising (NPD), a deep learning-based method for 3D point cloud denoising that estimates local reference planes using a neural network to project noisy points onto smooth surfaces, achieving state-of-the-art performance with lower variance than existing methods. NPD eliminates the need for iterative multi-projection or manual neighborhood selection by leveraging both local and global features for robust normal estimation and denoising in a single pass.

ABSTRACT

We present a neural-network-based architecture for 3D point cloud denoising called neural projection denoising (NPD). In our previous work, we proposed a two-stage denoising algorithm, which first estimates reference planes and follows by projecting noisy points to estimated reference planes. Since the estimated reference planes are inevitably noisy, multi-projection is applied to stabilize the denoising performance. NPD algorithm uses a neural network to estimate reference planes for points in noisy point clouds. With more accurate estimations of reference planes, we are able to achieve better denoising performances with only one-time projection. To the best of our knowledge, NPD is the first work to denoise 3D point clouds with deep learning techniques. To conduct the experiments, we sample 40000 point clouds from the 3D data in ShapeNet to train a network and sample 350 point clouds from the 3D data in ModelNet10 to test. Experimental results show that our algorithm can estimate normal vectors of points in noisy point clouds. Comparing to five competitive methods, the proposed algorithm achieves better denoising performance and produces much smaller variances.

Motivation & Objective

• To address the challenge of 3D point cloud denoising in the presence of noise, curvature variations, and varying sampling densities.
• To eliminate reliance on manually defined neighborhoods or eigendecomposition for local geometry estimation.
• To develop a deep learning-based method that estimates reference planes directly from noisy point clouds for accurate surface projection.
• To improve denoising performance by integrating both local and global contextual information in the network architecture.
• To achieve robust, stable denoising with minimal variance across diverse object categories and noise levels.

Proposed method

• A deep neural network is trained to predict normal vectors and plane intercepts (reference planes) for each point in a noisy 3D point cloud using local and global features.
• The network uses a PointNet-based architecture to extract features from local neighborhoods and global context, enabling joint estimation of local surface geometry.
• Noisy points are projected onto the estimated reference planes to produce denoised points, ensuring all output points lie on a smooth, continuous surface.
• A hybrid loss function combines MSE (point-to-point matching) and Chamfer distance (plane-to-plane matching), with a weighted parameter α to balance the two objectives.
• The model is trained on 40,000 point clouds from ShapeNet and tested on 350 point clouds from ModelNet10, with progressive training using increasing α values.
• Global feature injection significantly improves performance by providing structural context beyond local neighborhoods.

Experimental results

Research questions

• RQ1 Can a deep neural network effectively estimate reference planes for 3D point cloud denoising without relying on predefined neighborhoods?
• RQ2 Does integrating global context into local surface estimation improve robustness to noise intensity and curvature variation?
• RQ3 Can a single-projection denoising method outperform multi-projection baselines like WMP while reducing variance?
• RQ4 How does the balance between point-wise (MSE) and surface-wise (CD) matching affect denoising performance?
• RQ5 To what extent does the proposed method generalize across diverse object categories and noise levels?

Key findings

• NPD outperforms five competitive methods—BF, PDE, GBD, NLD, and WMP—in all seven categories (sofa, bed, toilet, etc.) based on both MSE and CD metrics.
• The proposed method achieves the lowest variance across all categories and evaluation metrics, demonstrating superior stability and robustness.
• For the sofa category, NPD achieves an MSE of 0.627 and CD of 0.7670, significantly lower than the 10.845 MSE and 10.4758 CD when using only local features.
• The inclusion of global features reduces MSE and CD errors by up to an order of magnitude compared to models using only local features.
• As α increases from 0.1 to 0.9, MSE decreases and CD increases, confirming that the loss function effectively balances point-wise and surface-wise matching.
• The model trained with increasing α values achieves optimal performance, showing that progressive training enhances generalization and denoising accuracy.