A Simple and Efficient Merge of Two Sparse 3D Models with Overlapped Images

Hanshin Lim, Soon-Heung Jung, Jeongil Seo

Abstract: One of the efficient approaches to generate a sparse 3D model is to incrementally merge the separately reconstructed 3D sub-models. This paper presents a simple and efficient method to merge two sparse 3D models having overlapped images. In the proposed method, first, a global similarity transform matrix between the two model is computed from the 3D points reprojected from the overlapped images. After that, the camera parameters of the overlapped images and the reprojected 3D points in one model are refined using the obtained global similarity transform matrix. A precise similarity transform matrix between the two 3D models is then estimated from the inlier 2D-3D correspondences to accurately align and merge the two 3D models. Experimental results show that the proposed method merge the two models more accurately compared to the existing merge method embedded in the state-of-the-art incremental Structure from Motion open software.