A review and evaluation of penalty functions for Semi-Global Matching

Multi-view matching presents significant advantages regarding accuracy, radiometric fidelity or occlusion handling; however, the limitation of having only two available views remains common in many real-life scenarios. The Semi-Global Matching (SGM) approach has been proved to be robust against complex scenes (from aerial images to those from autonomous systems) and radiometrically challenging images, providing accurate disparity maps (compared to global and local methods) and computational efficiency. Recently, an entire class of methods has appeared relying on SGM. In the present study, a review of methods which define an adaptive function for the P2 cost penalty for disparity discontinuities is presented. The evaluation methodology includes use of absolute intensity differences and the distance on Census transformed images as cost functions, and extensive parameter tuning with comparisons under the demanding Middlebury College 2006 stereo dataset, which involves radiometric changes in illumination/exposure. Results are reported and evaluated as regards the performance of such variations of SGM.