.png){kind=icon}
.png){kind=icon}
Return
Kanfilter: a simple and effective multi-module point cloud denoising model
.png){kind=icon}
PRE
AI
DOI：10.1007/s10586-024-05077-z.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The acquisition of point cloud data is often disrupted by noise, which affects the accuracy of downstream tasks such as surface analysis and precision measurement. This noise reduction process is known as point cloud denoising or filtering. Current deep learning-based methods face challenges, including limited model diversity, insufficient robustness in handling diverse data, and difficulties in ensuring that noisy points converge quickly and effectively to a smooth surface. To address these issues, this paper proposes a new model called KANFilter. KANFilter consists of multiple KANModules, each of which is composed of a KANBlock and a decoder to simulate the real denoising process in an integrated manner. Additionally, a new denoising target is designed to capture the relationships among intermediate filtering results during the training process. Experimental results show that this strategy significantly improves the quality of filtering results, achieving a maximum improvement of 39.3% in the CD (Chamfer Distance) metric at a 10K resolution compared to previous algorithms, and a 25% improvement in convergence speed relative to conventional algorithms.The source code can be found at: https://github.com/zhouxyly/KANFilter
Keywords:
Point cloud denoising
KAN
Graph network
Iterator denoising
Deep learning