Abstract:
Current stereo algorithms are capable to calculate accurate (as defined, e.g., by needs in vision-based driver assistance) dense disparity maps in real time. They have become the source of three-dimensional data for several indoor and outdoor applications. However, ground truth-based evaluation of such algorithms has been typically limited to data sets generated indoors in laboratories. In this paper we present a new approach to evaluate stereo algorithms using ground-truth over real world data sets. Ground truth is generated using range measurements acquired with a high-end laser range-finder. For evaluating as many points as possible in a given disparity map, we use two evaluation approaches: A direct comparison for those pixels with available range data, and a confidence measure for the remaining pixels.