A limitation of the state-of-art multi-view reconstruction algorithms is in their ability to handle scenes with very little texture or a lot of clutter. Texture-less scenes with clutter are traditionally difficult for dense multi-view stereo methods. Feature-based stereo algorithms, in particular, edgebased methods, are better suited for these scenes. The success of the edge-based reconstruction heavily depends on the amount of the prior information used. This paper introduces an algorithm that extends the well-known planesweep algorithm by incorporating spatial coherence of the scene. The method utilizes the geometric consistency derived from the continuity of three-dimensional edge points as a geometric constraint in addition to previously known multi-view image constraints. Experimental results on synthetic as well as real data are provided to demonstrate the efficacy and robustness of the proposed method.
Gang Li, Yakup Genc, Steven W. Zucker