Discarding wide baseline mismatches with global and local transformation consistency

Abstract

A novel method called global and local transformation consistency constraints, which combines the scale, orientation and spatial layout information of 'scale invariant feature transform' (SIFT) features, is proposed for discarding mismatches from given putative point correspondences. Experiments show that the proposed method can efficiently extract high-precision matches from low-precision putative SIFT matches for wide baseline image pairs, and outperforms or performs close to state-of-the-art approaches.Introduction: Local features are powerful tools for finding correspondences between wide baseline views of the same scenes. Some feature-based algorithms first establish putative correspondences, and then estimate the best global geometry relationship (such as homography) interpreting them. However, many well-known robust estimators (such as RANSAC [1]) perform poorly when the ratio of inliers is lower than 50% [2], while discarding mismatches before estimating this relationship yields important improvements, especially in the case where incorrect matches strongly outnumber the correct ones. Previous works (see e.g. [3,4]) for discarding mismatches mainly employ the geometrical and topological relationship among putative matches, but ignore the scale and orientation information of the potential feature pairs, which can express a similarity transformation.This Letter focuses on rejecting mismatches via evaluating the quality of each potential correspondence, which is measured by both global and local transformation consistency. To address the mismatches discarding problem, we divide the algorithm into two steps. First, using global constraint, we will be retaining a part of the matches of which the scale log-ratio and orientation difference are approximate to global scaling and rotation factor, respectively. Then, using local constraint, we will reject more incorrect matches from the first step, with a stricter constraint by requiring that neighbouring feature pairs have the similar transformation. Experiments show that the approach presented in this Letter improves the currently achieved wide baseline matching precision, with 10% fewer errors on most of the six well-known wide baseline image pairs, which were offered by Tuytellaars and Van Gool [5].

Robust key point descriptor for multi-spectral image matching

“…In this study, we focus on how to describe the com mon feature between multi-spectral images taken from the same scene, with the purpose of accurate matching, which is highly desirable in some applications, such as multi-spectral image fusion [14]. As to the wide-baseline situation [15], more schemes (e.g., topological clustering [15], global and local transformation consistency [16]) are needed and we will not go further with this situation in this study. SIFT key points are described by using the HCGEC, which is similar with histograms of oriented gradients (HOG) [17].…”