Acquiring photo-realistic 3D computer models of objects has been a very active research area with most important applications in Virtual Reality and Multimedia systems. This paper deals with a generalization of dense stereo reconstruction, which forms the basis for incremental fusing of stereo sequences to construct a 3-D model of underwater natural objects. The approach is fast and robust. Initialized with a set of robust feature matches over the left and right images, efficient left-to-right and right-to-left stereo matching are carried out by match propagation . Symmetry relation between the two stereo matchings is utilized to guide robust fusion. Where a large difference exists between the two dense stereo reconstructions, it is most probable that at least one estimate is erroneous, thus is rejected. Conversely, small differences are expected because of the discrete nature of the disparity values. When fused these two estimates, better results are expected. Experimental results demonstrate the efficacy of the proposed approach.