Multi-scale sparse feature point correspondence by graph cuts

This paper presents a global stereo sparse matching technique based on graph cut theory to obtain accurate correspondence of stereo point pair in vision measurement applications. First, in order to obtain accurate location of feature points, a new multi-scale corner detection algorithm is proposed where wavelet coefficients are used to determine feature points by calculating an auto-correlation matrix. A sparse graph is constructed based on the feature points according to the graph cut theory. Then, the feature point correspondence problem is transformed into a labeling problem in the sparse graph which can be solved by energy minimization. Multi-scale analysis is utilized to improve the precision of matching results. It was found that the use of sparse feature points in the construction of the graph can lead to both a simple graph structure and a reduced computational complexity. It was also found that node labeling in the graph can be performed using fewer disparity values instead of all disparity values. Our experimental results show that the new global stereo sparse matching technique can obtain more accurate results than the existing techniques.