[Paper Review] Non-Local Part-Aware Point Cloud Denoising

Chao Huang, Ruihui Li+2 more|arXiv (Cornell University)|Mar 14, 2020

3D Shape Modeling and Analysis41 references8 citations

TL;DR

This paper proposes a non-local part-aware deep learning framework for 3D point cloud denoising that leverages non-local self-similarity through a graph attention-based non-local learning unit (NLU) to capture semantically related features across the entire point cloud. By cascading NLUs and introducing a semantic part loss alongside surface reconstruction, the method achieves state-of-the-art performance on both synthetic and real-scanned data, significantly reducing noise while preserving fine geometric details and avoiding scattered artifacts.

ABSTRACT

This paper presents a novel non-local part-aware deep neural network to denoise point clouds by exploring the inherent non-local self-similarity in 3D objects and scenes. Different from existing works that explore small local patches, we design the non-local learning unit (NLU) customized with a graph attention module to adaptively capture non-local semantically-related features over the entire point cloud. To enhance the denoising performance, we cascade a series of NLUs to progressively distill the noise features from the noisy inputs. Further, besides the conventional surface reconstruction loss, we formulate a semantic part loss to regularize the predictions towards the relevant parts and enable denoising in a part-aware manner. Lastly, we performed extensive experiments to evaluate our method, both quantitatively and qualitatively, and demonstrate its superiority over the state-of-the-arts on both synthetic and real-scanned noisy inputs.

Motivation & Objective

• To address the limitations of local receptive field methods in point cloud denoising that fail to preserve fine geometric details and handle scattered noise.
• To exploit inherent non-local self-similarity in 3D point clouds by modeling long-range, semantically related features across distant regions.
• To improve denoising fidelity by introducing a part-aware loss that guides predictions toward relevant object parts.
• To achieve robust performance across diverse noise levels and data types, including real-scanned point clouds.
• To balance denoising performance and computational efficiency through ablation and hyperparameter analysis.

Proposed method

• Proposes a non-local learning unit (NLU) with a custom graph attention module to compute non-local, context-aware feature representations over the entire point cloud.
• Cascades multiple NLUs to progressively distill noise from the input, enabling hierarchical feature refinement.
• Introduces a semantic part loss that regularizes predictions to align with object parts, enhancing part-aware denoising.
• Combines the semantic part loss with standard surface reconstruction loss to jointly optimize for geometric fidelity and structural consistency.
• Employs a K-NN-based graph attention mechanism within the NLU to dynamically weigh contributions from distant, semantically similar points.
• Uses a multi-stage training strategy with ablation studies to validate component contributions and hyperparameter choices (e.g., K=16, 4 NLUs).

Experimental results

Research questions

• RQ1 Can non-local, part-aware feature learning improve point cloud denoising beyond local receptive field methods?
• RQ2 How does modeling long-range, semantically related features across distant regions affect denoising quality and geometric detail preservation?
• RQ3 To what extent does a part-aware loss improve the alignment of denoised points with underlying object structures?
• RQ4 What is the optimal number of NLUs and neighborhood size (K) for balancing performance and efficiency?
• RQ5 How does the proposed method compare to state-of-the-art approaches on both synthetic and real-scanned point clouds?

Key findings

• The full model with all components (NLU, part loss, cascaded NLUs) achieves the lowest Chamfer Distance (CD) of 2.31 × 10⁻⁴ on synthetic data with 1.5% noise, outperforming all baselines.
• Removing the graph attention unit leads to significantly higher CD values across all K settings, demonstrating its critical role in selecting relevant non-local features.
• Increasing K from small to 20 improves performance, with convergence observed at K=20, and K=16 chosen as a balance between accuracy and efficiency.
• Using five NLUs yields a slightly better CD (2.30 × 10⁻⁴) than four (2.31 × 10⁻⁴), but with higher parameter count (55.7K vs. 44.6K), justifying the use of four NLUs.
• The method reduces scattered artifacts and preserves sharp edges better than PointCleanNet and TotalDenoising, especially on real-scanned data like Paris-rue-Madame.
• The ablation study confirms that both the part-aware loss and the NLU module independently contribute to performance gains, with the largest improvement seen when both are added.