Multi-view Stereo Dense Reconstruction using SfM-based PMVS for Image-based Modeling (IBM)

Abstract

This paper intends to represent the effective Multiview stereo algorithm based on the image-based modeling (IBM). The system consists of three steps: patch initialization, expansion and filtering. In patch initialization, this approach takes the corresponding camera parameters together with sparse 3D points as inputs. The purpose is to introduce a framework that employs camera calibrating and 3D points from the results of structure-from-motion (SfM) method instead of Harris corner detector and Difference-of-Gaussians (DoG) in feature detection and matching step to initialize the patch. The patch expansion reconstructs at least one patch in every cell of the image. The patch expansion stage may contain outliers. Consequently, they remove the outlier patches in the filtering stage. These patch expansion and filtering are then iteratively implemented until getting the respectable and complete output. The experiments of our proposed framework on various datasets and comparisons between the other method are presented in this paper.