PointCartesian-Net: enhancing 3D coordinates for semantic segmentation of large-scale point clouds.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Collecting outdoor point cloud data is challenging due to complex algorithms and equipment costs.
• Limitations in data collection and point cloud characteristics hinder semantic segmentation development.

Purpose of the Study:

• To propose a novel neural network, PointCartesian-Net, for semantic segmentation of 3D point clouds.
• To utilize only 3D coordinates for semantic segmentation, reducing reliance on expensive equipment and complex data collection.

Main Methods:

• Encoding 3D coordinates to connect neighboring points and preserve geometric information.
• Employing dense and residual connections to expand receptive fields and aggregate multi-level, multi-scale semantic features.
• Introducing a 3D SENet module to learn channel-wise feature relations for improved weighting.

Main Results:

• Achieved 60.2% Mean Intersection-over-Union (mIoU) on the Semantic3D dataset.
• Obtained 89.1% overall accuracy on the large-scale Semantic3D benchmark.
• Demonstrated the feasibility and applicability of the proposed PointCartesian-Net.

Conclusions:

• PointCartesian-Net effectively performs semantic segmentation using only 3D coordinates.
• The network's architecture enhances feature extraction and contextual information aggregation.
• The results validate the potential of coordinate-based networks for 3D point cloud analysis.