Abstract: Dense matching is the key of dense 3D point cloud reconstruction based on stereo vision, and the elimination of mismatched points is an important means to ensure the matching accuracy. In this paper, results of sparse matching are used as seeds, and quasi-dense matching is obtained by matching propagation in the neighborhood of the seeds. To reduce the time consumption of quasi-dense matching measurement while ensuring the matching accuracy, the local shading consistency based on Retinex theory is proposed, which is combined with the confidence of the brightness difference to eliminate the impossible matching in advance. Then, quasi-dense matching with similarity measurement is carried out. Experimental results show that the algorithm achieves smaller average disparity error on the training set of Middlebury 2014; On the scan114 of DTU dataset and the captured image data, the algorithm can effectively reduce the quasi-dense matching time and the reprojection error of 3D reconstruction.