Point Cloud Noise and Outlier Removal with Locally Adaptive Scale.

This paper introduced a simple and effective algorithm to remove the noise and outliers in point sets generated by multi-view stereo methods. Our main idea is to discard the points that are geometrically or photometrically inconsistent with its neighbors in 3D space using the input images and corresponding depth maps. We attach a scale value to each point reflecting the influence to the adjacent area of the point and define a geometric consistency function and a photometric consistency function for the point. We employ a very efficient method to find the neighbors of a point using projection. The consistency functions are related to the normal and scale of the neighbors of points. Our algorithm is locally adaptive, feature preserving and easy to implement for massive parallelism. It performs robustly with a variety of noise and outliers in our experiments.