Deterministic Point Cloud Diffusion for Denoising.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Diffusion models excel at image restoration by refining noisy data.
• Applying diffusion models to 3D point cloud denoising faces challenges due to structured displacements and noise complexity.
• Existing methods struggle with geometric relationships and surface recovery in point cloud denoising.

Purpose of the Study:

• To develop a deterministic noise-free diffusion framework for point cloud denoising.
• To address the limitations of traditional Gaussian noise diffusion in 3D domains.
• To achieve faithful surface recovery and mitigate common denoising artifacts.

Main Methods:

• A two-phase residual diffusion process is proposed for point cloud denoising.
• The forward phase injects directional residuals into clean surfaces, creating a degradation trajectory.
• The reverse phase utilizes a U-Net-based network to estimate and remove residuals, recovering the surface.

Main Results:

• The proposed method achieves state-of-the-art performance on synthetic and real-world datasets.
• Quantitative metrics and visual quality demonstrate superior denoising capabilities.
• The framework successfully mitigates over-smoothing and under-smoothing artifacts.

Conclusions:

• The deterministic noise-free diffusion framework offers a robust solution for point cloud denoising.
• The two-phase residual diffusion process effectively handles structured displacements and geometric reasoning.
• This approach enables faithful 3D surface recovery with high fidelity.