Learning Implicit Fields for Point Cloud Filtering.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Noisy point clouds are a common challenge in 3D data acquisition.
• Traditional filtering methods struggle with preserving sharp features and require extensive parameter tuning.
• Existing data-driven methods often blur details or lead to uneven point distribution.

Purpose of the Study:

• To develop a novel data-driven method for denoising 3D point clouds.
• To overcome limitations of existing filtering algorithms, particularly in feature preservation.
• To leverage implicit field representations for robust point cloud cleaning.

Main Methods:

• Exploration of implicit fields and predicted signed distance fields (SDFs).
• A novel encoder-decoder architecture for processing aligned local point cloud patches.
• Separating point movement direction and distance prediction using SDFs and gradient descent.

Main Results:

• The proposed method achieves feature-preserving point cloud filtering without explicit normal estimation.
• Visual and quantitative experiments show superior performance compared to state-of-the-art methods.
• The technique outperforms traditional and learning-based position-based denoising methods.

Conclusions:

• The implicit field-based approach offers a robust solution for noisy point cloud denoising.
• This method effectively preserves geometric details and sharp features.
• It presents a significant advancement in 3D geometry processing applications.