Globally consistent correspondence of multiple feature sets using proximal Gauss–Seidel relaxation. (March 2016)

• Record Type: Journal Article
• Authors: Yu, Jin-Gang; Xia, Gui-Song; Samal, Ashok; Tian, Jinwen
• Is Part Of: Pattern recognition. Volume 51(2016:Mar.)
• Page Start: 255
• Page End: 267
• Publication Date: 2016-03
• Subjects: Feature correspondence -- Multiple feature set correspondence -- Permutation matrix -- Convex relaxation -- Proximal Gauss–Seidel method -- Graph matching; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2015.09.029
• Languages: English
• ISSNs: 0031-3203

Abstract

Abstract: Feature correspondence between two or more images is a fundamental problem towards many computer vision applications. The case of correspondence between two images has been intensively studied, however, few works so far have been concerned with multi-image correspondence. In this paper, we address the problem of establishing a globally consistent correspondence among multiple (more than two) feature sets given the pairwise feature affinity information. Our main contribution is to propose a novel optimization framework for solving this problem based on the so-called Proximal Gauss–Seidel Relaxation (PGSR). The proposed method is distinguished from previous works mainly in three aspects: (1) it is more robust to noise and outliers; (2) its solution is based on convex relaxation and the principled PGSR method, which in general has convergence guarantee; (3) the scale of the problem in our method is linear with respect to the number of feature sets, making it computationally practical to be used in real-world applications. Experimental results both synthetic and real image datasets have demonstrated the effectiveness and superiority of the proposed method. Abstract : Highlights: We study finding globally consistent correspondence from multiple feature sets. A novel method is proposed based on Proximal Gauss–Seidel Relaxation (PGSR). PGSR is more robust to noise, outliers and deformation than competitors. The optimization has general convergence guarantee. The scale of our formulation is linear w.r.t. the number of feature sets.