Pointfilter: Point Cloud Filtering via Encoder-Decoder Modeling.

Area of Science:

• Computer Vision
• Geometric Modeling
• Machine Learning

Background:

• Point cloud filtering is crucial for geometry processing.
• Existing methods struggle with preserving sharp features and require extensive parameter tuning.

Purpose of the Study:

• To develop an automated and robust deep learning approach for point cloud filtering.
• To address the limitations of existing methods in noise removal and sharp feature preservation.

Main Methods:

• A novel point-wise deep learning architecture with an encoder-decoder structure was employed.
• The encoder learns latent representations from point neighborhoods.
• The decoder predicts displacement vectors to generate clean point coordinates.

Main Results:

• The proposed method automatically filters noisy point clouds, preserving sharp features.
• Achieved superior performance compared to state-of-the-art deep learning techniques.
• Demonstrated high visual quality and improved quantitative error metrics.

Conclusions:

• The novel deep learning approach offers an effective solution for point cloud filtering.
• It robustly removes noise and preserves critical sharp features.
• The method reduces the need for manual parameter tuning, simplifying the process.