Abstract: Stereo vision is one of the critical techniques in computer vision, which determines conjugate pairs in stereo image and measures height of pixel robustly and accurately. But it can only produce a sparse set of surface depths due to the difficulty of identifying corresponding feature points in the region where gray level varies gradually. The integration of shape from shading and stereo vision is considered to be the possible outcome of the investigation of solving this problem. In this paper, the ill-posed nature of shape from shading is explained, and is applied to explicate the unreliability, instability, and limitation of the current SFS algorithms. So these algorithms are impracticable to be put into the integrated system directly. The methods of integrating the two vision modules are introduced in four categories, namely, mergence, adjustment, combination and cooperation. And it is illustrated that, the reasonable implementation of an integrated system should involve models of each vision module and models of their interaction. This could not only make it possible to overcome the ambiguity of shape from shading with the supplementary information supplied by stereo module, bur also benefit the performance of the whole vision system. Keywords Shape from shading, Stereo vision, Surface orientation, Reflective albedo

摘要： 立体视觉(Stereo Vision)方法是目前利用图象数据获取物体三维信息的主要方法之一。但该方法在图象灰度变化较缓慢的区域，由于难以准确地进行图象间的象素配对，而严重影响了它的效果。利用从明暗重构物体三维表面形状(Shape from Shading,简称SFS)的方法与该方法相结合，是改善重构结果的主要途径之一。文章通过分析SFS问题本身的不适定性，揭示了目前几类主要的SFS算法在可靠性、稳定性、局限性以及实用性方面所存在的问题，并在此基础上，简要地介绍了四类SFS与立体视觉方法相结合的形式，说明了通过利用立体视觉为SFS补充辅助的信息来消除SFS问题的不适定性，并对过去SFS的实现方法进行有效的改进，是提高集成系统准确性的关键。