Related Experiment Video
Updated: May 28, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
353
RotInv-PCT: Rotation-Invariant Point Cloud Transformer via feature separation and aggregation.
Cheng He1, Zhenjie Zhao1, Xuebo Zhang1
1Institute of Robotics and Automatic Information System, College of Artificial Intelligence, Nankai University, Tianjin, China; Tianjin Key Laboratory of Intelligent Robotics, Nankai University, Tianjin, China.
Summary
This study introduces a Rotation-Invariant Point Cloud Transformer (RotInv-PCT) that enhances neural network performance by incorporating relative pose features. RotInv-PCT achieves superior rotation invariance for point cloud processing tasks, outperforming existing methods.
Area of Science:
- Computer Vision
- Machine Learning
- Geometric Deep Learning
Background:
- Point clouds are increasingly utilized, driving demand for robust neural network processing.
- Rotation invariance is critical for consistent point cloud analysis but current methods using local coordinate systems have limitations.
- Existing approaches often overlook relative pose information between local point cloud structures, impacting performance.
Purpose of the Study:
- To develop a novel neural network architecture for rotation-invariant point cloud processing.
- To improve the performance of point cloud analysis tasks by explicitly modeling relative pose relationships.
- To achieve state-of-the-art results in point cloud classification, segmentation, and other related tasks.
Main Methods:
- Proposed Rotation-Invariant Point Cloud Transformer (RotInv-PCT) utilizing Local Reference Frames (LRFs) for shape features.
- Introduced a Feature Aggregation Transformer (FAT) to fuse rotation-invariant shape and pose features for global representation.
- Employed hierarchical random downsampling for efficient processing of large-scale point clouds.
Main Results:
- RotInv-PCT demonstrated superior performance across various benchmarks including ScanObjectNN, ModelNet40, ShapeNet, S3DIS, and KITTI.
- Achieved a 2% improvement in real-world point cloud classification compared to the strongest baseline.
- Significantly improved semantic segmentation on S3DIS by 10% and enabled the first rotation-invariant semantic segmentation on KITTI.
Conclusions:
- The proposed RotInv-PCT effectively addresses limitations of existing methods by incorporating relative pose information.
- The novel approach provides provably rotation-invariant features and achieves state-of-the-art results in diverse point cloud tasks.
- RotInv-PCT offers a significant advancement in handling and analyzing 3D point cloud data with enhanced robustness and accuracy.
