A method for using random forests to estimate image correspondences is proposed. The method does not rely on the availability of manual label annotations. Labels for training are obtained via the use of supervoxels. The efficient method is effective at providing an estimate of image correspondences.