Application of balanced neural tree for classifying tentative matches in stereo vision

Here, we present a new application of the supervised learning based classifier in stereo matching. In particular, the chosen classifier is a balanced neural tree. The tentative matches are obtained using the speeded-up robust feature (SURF) matching. The feature vector corresponding to each tentative match is formed based on a similarity measure between SURF descriptors and their neighborhoods in the two stereo images, and these feature vectors are classified into inlier or outlier classes. Further, accuracy of the obtained results have been evaluated in terms of stereovision applications such as in the estimation of the homography and rectification matrices between two stereo images. The experiments based on these stereo estimates show the applicability of the proposed application in the stereo vision.