Robust point correspondence by improved proximity matrix

Abstract

This paper proposed a new improved singular value decomposition method to achieve high accuracy and much more number of correct point correspondences between uncalibrated images with large scene variations. The proposed matching method is based on singular value decomposition and Sift feature descriptor. The proximity matrix for decomposition is redefined to improve the performance of robustness and reliability. Firstly the distance of Sift descriptors is introduced in the proximity matrix to replace spatial distance. Furthermore illumination invariant normalized cross correlation, that simultaneously includes scale and dominant orientation of the feature points, is used as similarity measure to strengthen proximity matrix. Thus, the element in proximity matrix is invariant to scale, rotation, and light changes. Experimental results show that the improved method can be used for point correspondence with severe wide baseline variations and provide evidence of better performance with respect to other popular algorithms.