Learning Semantic Segmentation of Large-Scale Point Clouds With Random Sampling.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Existing semantic segmentation methods for 3D point clouds are computationally intensive.
• Current approaches struggle with large-scale datasets due to sampling or pre/post-processing limitations.

Purpose of the Study:

• To introduce RandLA-Net, an efficient and lightweight neural architecture for direct semantic segmentation of large-scale 3D point clouds.
• To overcome the limitations of existing methods in processing extensive 3D data.

Main Methods:

• Utilized random point sampling as a core component for efficiency.
• Introduced a novel local feature aggregation module to preserve geometric details.
• Developed a lightweight neural architecture for direct per-point semantic inference.

Main Results:

• RandLA-Net processes 1 million points up to 200x faster than existing methods.
• Achieved state-of-the-art semantic segmentation performance across five large-scale datasets.
• Demonstrated significant computational and memory efficiency.

Conclusions:

• RandLA-Net provides an effective solution for efficient semantic segmentation of large-scale 3D point clouds.
• The proposed method balances efficiency with high performance, preserving crucial geometric details.
• This architecture enables practical application on massive 3D datasets.