Guided normal filter for 3D point clouds

Point clouds have been attracting more and more attention due to the advancement of 3D sensors. However, the raw point clouds acquired suffer inevitably from noise, which challenges their applications in 3D computer vision. In order to address this problem, we propose a novel feature-preserving filtering framework, termed Guided Normal Point Cloud Filter. First, we perform initial normal estimation using improved Principal Component Analysis algorithm. Then, a well-designed point normal filter based on locally linear model is proposed, which uses the estimated normal field as guidance. Finally, according to the adjusted normal field, we treat the point positions update problem as a least-squares issue solved by stochastic gradient decent optimizer. Quantitative and qualitative experimental results on several point cloud models show the effectiveness of our proposed algorithm, which can provide a much better trade-off between filtering performance and computational efficiency.