Stereo matching has received a lot of attention from the computer vision community, thanks to its wide range of applications. Despite of the large variety of algorithms that have been proposed so far, it is not trivial to select suitable algorithms for the construction of practical systems. One of the main problems is that many algorithms lack sufficient robustness when employed in various operational conditions. This problem is due to the fact that most of the proposed methods in the literature are usually tested and tuned to perform well on one specific dataset. To alleviate this problem, an extensive evaluation in terms of accuracy and robustness of state-of-the-art stereo matching algorithms is presented. Three datasets (Middlebury, KITTI, and MPEG FTV) representing different operational conditions are employed. Based on the analysis, improvements over existing algorithms have been proposed. The experimental results show that our improved versions of cross-based and cost volume filtering algorithms outperform the original versions with large margins on Middlebury and KITTI datasets. In addition, the latter of the two proposed algorithms ranks itself among the best local stereo matching approaches on the KITTI benchmark. Under evaluations using specific settings for depth-image-based-rendering applications, our improved belief propagation algorithm is less complex than MPEG's FTV depth estimation reference software (DERS), while yielding similar depth estimation performance. Finally, several conclusions on stereo matching algorithms are also presented.