We present an alternative approach to address correspondence in computational stereo. Accurate correspondence between two 2D images is required to determine depth, which leads to 3D model generation. To achieve this, optimized noise reduction method and accurate edge detection are required for preprocessing, followed by extracting the contours from the edge detection output and detecting the correspondence for each of them. This detection follows a hierarchal method we termed level-based correspondence. The results show improvements in detection speed and accuracy, compared to other local correspondence methods. © 2010 IEEE.