Abstract: A novel multi-stage quasi-dense matching algorithm for wide base-line stereo images is introduced based on SIFT and dual constraints of epipolar geometry and homographic mapping. The proposed algorithm includes following three stages: (1) the optimal SIFT features with good spatial distribution and large information content are first selected, and matched by using the least squares matching method, then the fundamental and homographic matrix can be estimated by using these initial sparse correspondences with higher precision; (2) For the other SIFT features, affine transformation parameters between matching windows are iterative optimized by using the slope angle of correspondent epipolar lines and scale information of SIFT features, and affine invariant feature descriptors are extracted from the corrected matching windows, then correspondences can be determined by Euclid distance and dual constraint information; (3) Considering the lower repeatability rate of feature detection for wide base-line stereo images, for the unmatched points extracted from left and right images of stereo pairs, matching can be carried out by adopting two-way search strategy from left to right image or from right to left image based on the rapid SSD similarity cost function and affine rectified dual constraints region, and the least squares curve surface fitting weighted by Gaussian-distance algorithm is adopted to improve the precision of matching results. Test results using practical wide base-line image pairs indicate the effectiveness of the proposed algorithm and can provide reliable dense or quasi-dense matching points for subsequent 3D reconstruction.