Core Concepts
VecKM is a novel local point cloud geometry encoder that is descriptive, efficient, and robust to noise. It achieves this by vectorizing a kernel mixture representation of the local point cloud, which is proved to be reconstructive and isometric to the original local shape.
Abstract
The paper proposes VecKM, a novel local point cloud geometry encoder that is highly efficient and robust to noise.
Key highlights:
VecKM encodes the local point cloud by vectorizing a kernel mixture representation, which is proved to be reconstructive and isometric to the original local shape.
VecKM is the only existing local geometry encoder that costs linear time and space (O(nd)), achieved through its unique factorizable property.
Extensive experiments show that VecKM outperforms existing encoders in terms of accuracy, speed, and robustness to noise across various point cloud tasks, including normal estimation, classification, part segmentation, and semantic segmentation.
VecKM can be seamlessly integrated into deep point cloud architectures, significantly improving their efficiency while maintaining or improving their performance.
The paper provides a solid theoretical foundation for VecKM's descriptiveness, efficiency, and noise robustness. It also presents detailed experiments demonstrating VecKM's superior performance compared to existing local geometry encoders.
Stats
VecKM is 100x faster than existing encoders in normal estimation tasks.
VecKM achieves up to 9.5x faster inference time compared to baseline point cloud architectures while maintaining or improving their performance.
Quotes
"VecKM is the only existing local geometry encoder that costs linear time and space (O(nd))."
"Extensive experiments show that VecKM outperforms existing encoders in terms of accuracy, speed, and robustness to noise across various point cloud tasks."
"VecKM can be seamlessly integrated into deep point cloud architectures, significantly improving their efficiency while maintaining or improving their performance."