Binocular Vision has been considered to be a method of three dimensional measurement, and this leads to lack of discussing how to decide the size of region that apply physical restriction, the reliability of disparities. These missing issues are quite important for robots when they use it as a method of percepting depth information. In this paper, we propose an adaptive algorithm to compute reliable disparities fast. It is composed of two processes. One is the process that compute disparities with the uniform radius of the correlation window. It can compute disparities with high reliability by eliminating false disparities sequencially. Another is the process that compute disparities with the adaptive radius of the correlation window. Since this process automaticaly selects the minimum radius of the correlation window in which reliable disparities can be detected, fast computation of disparities is realizable. It is shown that the algorithm of Adaptive Disparity Detection is effective for various images.