ResSANet: Learning Geometric Information for Point Cloud Processing

Area of Science:

• Computer Vision
• Geometric Deep Learning
• 3D Point Cloud Processing

Background:

• Point clouds offer rich geometric data crucial for robotics and autonomous driving.
• Existing methods struggle to extract sufficient geometric features from raw point clouds, limiting downstream task performance.
• Accurate feature extraction is essential for point cloud classification, shape retrieval, and part segmentation.

Purpose of the Study:

• To propose a novel convolution method for enhanced geometric feature extraction from 3D point clouds.
• To introduce ResSANet, a new convolution neural network designed to learn hierarchical geometry information.
• To improve the performance of point cloud processing tasks through advanced feature representation.

Main Methods:

• A novel convolution based on geometric primitives is developed to represent point cloud shapes.
• A new convolution neural network, ResSANet, is constructed using geometric primitives.
• Two modules, Res-SA and Res-SA-2, are devised for multi-level feature fusion within ResSANet.

Main Results:

• Achieved 93.2% classification accuracy on the ModelNet40 dataset.
• Reached 87.4% accuracy in shape retrieval tasks.
• Obtained 83.3% (class mIoU) and 85.3% (instance mIoU) accuracy in part segmentation on the ShapeNet dataset.
• The network features minimal depth and only 1.04 M parameters.

Conclusions:

• ResSANet effectively extracts hidden geometric features from point clouds.
• The proposed method demonstrates superior performance compared to state-of-the-art approaches.
• ResSANet offers an efficient yet powerful solution for 3D point cloud analysis.