Improving stereovision matching through supervised learning

Abstract

Most classical local stereovision matching algorithms use features representing objects in both images and compute the minimum difference attribute values. We have verified that the differences in attributes for the true matches cluster in a cloud around a centre. The correspondence is established on the basis of the minimum squared Mahalanobis distance between the difference of the attributes for a current pair of features and the cluster centre (similarity constraint). We introduce a new supervised learning strategy derived from the Learning Vector Quantization (LVQ) approach to get the best cluster centre. Additionally, we obtain the contribution or specific weight of each attribute for matching. We improve the learning law introducing a variable learning rate. The supervised learning and the improved learning law are the most important findings, which are justified by the computed better results compared with classical local stereovision matching methods without learning and with other learning strategies. The method is illustrated with 47 pairs of stereo images.