Abstract :
Stereo-matching is an indispensable process of dense 3D information extraction for a wide range of applications. Relevant methods rely on cost functions and optimization algorithms for estimating accurate disparities. This work analyses a novel cost for stereo matching under radiometric differences in the stereo-pair, which is based on a modification of the widely used census transformation. It is proposed to define the census on image x and y gradients. The modified census (MC) on gradients is evaluated as an independent matching cost in the presence of severe radiometric differences. For this, the original and the modified census transformation (CT) are implemented in three different aggregation schemes, namely fixed rectangular windows, adaptive cross-based support regions and semi-global matching. It is shown that the MC can provide better results in the cases of local radiometric differences, such as different illumination conditions. Thus, this approach can extend the inherent capability of the original CT to address global monotonic radiometric differences.
