While many algorithms for computing stereo correspondence have been proposed, there has been very little work on experimentally evaluating algorithm performance, especially using real (rather than synthetic) imagery. In this paper we propose an experimental comparison of several different stereo algorithms. We use real imagery, and explore two different methodologies, with different strengths and weaknesses. Our first methodology is based upon manual computation of dense ground truth. Here we make use of a two stereo pairs: one of these, from the University of Tsukuba, contains mostly fronto-parallel surfaces; while the other, which we built, is a simple scene with a slanted surface. Our second methodology uses the notion of prediction error, which is the ability of a disparity map to predict an (unseen) third image, taken from a known camera position with respect to the input pair. We present results for both correlation-style stereo algorithms and techniques based on global methods such as energy minimization. Our experiments suggest that the two methodologies give qualitatively consistent results. Source images and additional materials, such as the implementations of various algorithms, are available on the web from http://www.research.microsoft.com/szeliski.