PyraPVConv: Efficient 3D Point Cloud Perception with Pyramid Voxel Convolution and Sharable Attention.

Area of Science:

• Computer Vision
• Deep Learning
• 3D Data Processing

Background:

• Deep learning models for 3D point cloud perception are crucial.
• Existing Point-Voxel Convolution (PVConv) methods have limitations in performance.
• There is a need for more efficient and effective 3D convolution techniques.

Purpose of the Study:

• To propose a novel Pyramid Point-Voxel Convolution (PyraPVConv) block.
• To enhance feature extraction and fusion for 3D point cloud data.
• To improve the performance and efficiency of deep learning models in 3D perception tasks.

Main Methods:

• Introduced a voxel pyramid module for efficient, multi-scale feature extraction.
• Implemented a sharable attention module for feature aggregation and complexity reduction.
• Stacked PyraPVConv blocks to construct deep learning networks.

Main Results:

• PyraPVConv networks demonstrate superior performance on indoor scene segmentation, object part segmentation, and 3D object detection.
• The proposed method achieves efficiency in GPU memory consumption and computational complexity.
• Results validate the effectiveness of the novel PyraPVConv block.

Conclusions:

• PyraPVConv offers significant improvements over existing methods for 3D point cloud perception.
• The proposed block enhances feature representation and fusion capabilities.
• PyraPVConv is an efficient and effective solution for various 3D perception tasks.