Spherical Interpolated Convolutional Network With Distance-Feature Density for 3-D Semantic Segmentation of Point...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Semantic segmentation of 3D point clouds is crucial for robot environment perception.
• Traditional 3D convolution struggles with unstructured point cloud data.
• Existing interpolation methods have limitations in accurately weighting point cloud features.

Purpose of the Study:

• To develop a more effective feature extraction method for 3D point cloud semantic segmentation.
• To address the limitations of traditional 3D convolution operators on unstructured point cloud data.
• To improve the accuracy and efficiency of semantic segmentation networks for robotic applications.

Main Methods:

• A novel spherical interpolated convolution operator is proposed to replace traditional grid-shaped 3D convolution.
• A self-learned distance-feature density method is introduced, combining distance and feature correlation for interpolation.
• The proposed method is integrated into a spherical interpolated convolution network for feature extraction.

Main Results:

• The proposed spherical interpolated convolution network demonstrates effectiveness in 3D point cloud semantic segmentation.
• Experiments on ScanNet and Paris-Lille-3D datasets show good performance.
• Comparison experiments confirm improved accuracy and reduced network parameters compared to traditional 3D convolution.

Conclusions:

• The spherical interpolated convolution operator provides a more rational and effective approach to 3D point cloud feature extraction.
• The self-learned distance-feature density enhances the interpolation process for better segmentation results.
• The proposed method offers a significant advancement for semantic segmentation in robotic perception systems.