Abstract
In this paper, we show that we can improve accuracies of 3-D reconstructions with uncalibrated stereo by classifying correspondences between two images. After obtaining initial correspondences by an automatic matching program, we classify the correspondences into inliers and outliers in a multi-dimensional feature space. For doing this, we introduce four quantities with respect to a corresponding pair and adopt EM algorithm and one-class SVM with a special kernel fitted to their characteristics. By real image examples, we show that we can improve the accuracies of 3-D reconstructions by classifying correspondences.