Abstract: Dense stereo image matching is a key technique to find correspondences through fixed matching windows and then compute 3D points through the triangulation measure. Its advantages of low cost, high point density and large measure area have fueled several smart 3D applications. However, due to occlusions in building boundaries, the fixed matching window often fattens the boundaries to a certain extend, which greatly reduces the matching accuracy in building boundaries. To achieve higher-accuracy matching result in building boundaries, this paper proposes a global building boundary refinement method based on line features, which firstly extracts line features in disparity/elevation jumps as building boundaries and then globally refines these boundaries under the basic assumption that pixels with similar intensities should have the similar disparities. The main contribution of the algorithm is to formulate the building boundary refinement problem as the optimization of a new global energy function, which is able to sharpen boundaries as well as keep details around boundaries. Compared with some state-of-the-art boundary refinement algorithms (e.g. local sharpen operators and the plane based boundary refinement), the proposed method is capable of addressing the issues they met, i.e. the incapacity to correct large errors or the over smoothness around boundaries. Experiments on aerial datasets and satellite datasets show that the proposed method is superior to two other popular boundary sharper operators and a state-of-the-art plane based boundary refinement method, and can efficiently reduce the errors in boundaries. Therefore, our method can be applied in several 3D applications, e.g. virtual reality, smart city, and building extraction.

Key words: line segment detector, graph cuts optimization, dense stereo image matching, building boundary refinement, guided image filtering