Better Correspondence by Registration

Accurate image correspondence is crucial for estimating multiple-view geometry. In this paper, we present a registration-based method for improving accuracy of the image correspondences. We apply the method to fundamental matrix estimation under practical situations where there are both erroneous matches (outliers) and small feature location errors. Our registration-based method can correct feature locational error to less than 0.1 pixel, remedying localization inaccuracy due to feature detectors. Moreover, we carefully examine feature similarity based on their post-alignment appearance, providing a more reasonable prior for subsequent outlier detection. Experiments show that we can improve feature localization accuracy of the MSER feature detector, which recovers the most accurate feature localization as reported in a recent study by Haja and others. As a result of applying our method, we recover the fundamental matrix with better accuracy and more efficiency.