Self-adaptive normal estimation and position adjustment for MVS reconstruction

Abstract

Generating a polygonal mesh model from the point cloud is a critical step of many state-of-art MVS reconstruction algorithms, and influences the accuracy and visual quality of the final results significantly. The normal estimation and position adjustment of each point is required for this procedure. We present a mathematical analysis of the normal estimation approach, and propose two hypotheses to determinate the accuracy and smoothness of the points in a local region. A multi-scale strategy is implemented to obtain a proper scale for each point. Then the according normal is calculated by PCA on this scale, and the positions can be optimized by combining the accurate neighboring normals. A 2D toy example proves that the proposed approach can adjust the noisy point to the right surface while preserving details. At last we show that our method can improve the quality of mesh models for real MVS reconstruction tasks.