Efficient and accurate 3D modeling based on a novel local feature descriptor

Abstract

Registration is a key step in 3D modeling. In this paper, we propose an efficient and accurate 3D modeling algorithm composed of pairwise registration and multi-view registration. In pairwise registration, we propose a novel local descriptor named divisional local feature statistics (DLFS) which is generated by first dividing a local space into several partitions along projected radial direction, and then performing the statistics of one spatial and three geometrical attributes on each partition. For improving the compactness of DLFS, a principal component analysis (PCA) technique is used to compress it. Based on the compressed DLFS descriptor together with a game theoretic matching technique and two variants of ICP, the pairwise registration is efficiently and accurately performed. On this basis, a multi-view registration is performed by combining shape growing based registration technique and simultaneous registration method. In this process, a correspondence transition technique is proposed for efficiently and accurately estimating the overlap ratio between any two inputting scans. Extensive experiments are conducted to verify the performance of our algorithms. The results show that the DLFS descriptor has strong robustness, high descriptiveness and efficiency. The results also show that the proposed 3D modeling algorithm is very efficient and accurate.