Rotation-Invariant Point Cloud Representation for 3-D Model Recognition.

Area of Science:

• Computer Vision
• 3-D Data Analysis
• Machine Learning

Background:

• Point clouds are a key 3-D data modality in computer vision.
• Current methods often lack rotation invariance, limiting 3-D object recognition performance.
• Achieving rotation invariance is essential for robust point cloud representation.

Purpose of the Study:

• To develop a rotation-invariant representation for 3-D point clouds.
• To improve the performance of deep learning models for 3-D object recognition.
• To address the limitations of permutation-invariant approaches.

Main Methods:

• Introduced an equivalence relation under the rotation group SO(3).
• Located point cloud representations in a homogeneous space invariant to rotation.
• Integrated a convolutional network to parameterize SO(3) and capture all 3-D rotations.
• Selected optimal rotation as the best representation and minimized the problem on SO(3) using its geometric structure.

Main Results:

• The proposed network is flexibly incorporated into existing point cloud frameworks.
• Demonstrated rotation invariance by combining the approach with two deep models.
• Evaluated performance on the ModelNet40 and ModelNet10 datasets.
• Showed improved performance of existing deep models with arbitrary rotations.

Conclusions:

• The developed strategy effectively achieves rotation invariance for 3-D point cloud representations.
• The approach enhances the robustness and accuracy of 3-D object recognition systems.
• This method offers a significant advancement in handling 3-D data in computer vision.