Brute-force dense matching is usually not satisfactory because the same search range is used for the entire image, yielding potentially many false matches. In this paper, we propose a progressive scheme for stereo matching which uses two fundamental concepts: the disparity gradient limit principle and the least commitment strategy. The first states that the disparity should vary smoothly almost everywhere, and the disparity gradient should not exceed a certain limit. The second states that we should first select only the most reliable matches and therefore postpone unreliable decisions until enough confidence is accumulated. Our technique starts with a few reliable point matches obtained automatically via feature correspondence or through user input. New matches are progressively added during an iterative matching process. At each stage, the current reliable matches constrain the search range for their neighbors according to the disparity gradient limit, thereby reducing potential matching ambiguities of those neighbors. Only unambiguous matches are selected and added to the set of reliable matches in accordance with the least commitment strategy. In addition, a correlation match measure that allows rotation of the match template is used to provide a more robust estimate. The entire process is cast within a Bayesian inference framework. Experimental results illustrate the robustness of our proposed dense stereo matching approach.