For the purposes of three-dimensional modeling of buildings, a new stereo image matching algorithm based on the gray-scale projection is advanced considering the characteristics of the digital close-range photogrammetry. The algorithm extends the one-dimensional gray-scale projection from two directions to four directions. The gray-scale projections in the column, row, major diagonal and secondary diagonal directions are calculated simultaneously. The weighted average of similarity measure of these four one-dimensional projection vectors between the left image and right image is computed for the final matching decision. Homonymous epipolar lines are found by using the relative orientation linear transformation (RLT). To decrease the searching scope and improve the matching accuracy, an improved dynamic bounding approach is proposed to control the searching scope of sequential points in the point list to be matched along the homonymy epipolar-line. The new algorithm is applied to actual stereo images of a building taken by a digital close-range photogrammetric system and the experimental result shows that the algorithm has a higher matching speed and matching accuracy than the pyramid image matching algorithm.