3D point cloud denoising and normal estimation for 3D surface reconstruction

Denoising numerous large scale noise and preserving fine features simultaneously remains a challenge to point-cloud-related multiple view stereo (MVS) reconstruction approaches. The proposed algorithm reuses the sparse point cloud which is often discarded after the structure form motion (SfM) procedure in image based modeling to guide the dense point cloud denoising. Furthermore, the utilization of the octree division provides an efficient and simple denoising mechanism. Experiments show that the proposed method successfully removes the large scale noise points and presents a satisfactory denoising result with detailed information preserved. In addition, the normal of each point can be estimated fast and accurately as a by-product of the denoising algorithm.