Sift feature point selection by using image segmentation
Yuji Nakashima, Yoshimitsu Kuroki · 2017
The purpose of this study is to improve corresponding points detection accuracy. Corresponding points between multiple images taking same objects are determined by using feature points of the images. SIFT is a well-known method to find corresponding points; however, SIFT feature points with a similar feature vectors cause erroneous correspondents. This paper selects SIFT feature points on only foreground region of images because corresponding points between detected foreground and background regions seem to be erroneous. This paper proposes to segments an image automatically by using Graph Cut. Our method also selects valid corresponding points using estimated fundamental matrix, and corresponding points are eliminated when the points do not satisfy the epipolar constraint. Experimental results show that the proposed method increased correct matching rate from 49% to 74% by reducing erroneous correspondents.