Abstract: Textures of outdoor buildings are generally repetitive and locally insufficient. Previous quasi-dense matching methods do not work well on such wide baseline images because the obtained initial seeds are not enough and have matching ambiguity when propagating. This paper proposes an efficient quasi-dense matching algorithm for wide baseline images of building scene, of which MSERDoG(maximally stable extremal regions on difference of Gaussian space) is given to obtain more initial seed matches, and then affine propagation with adaptive support- weight score is used to have better quasi-dense matches. Experiments demonstrate that the proposed algorithm is efficient and satisfactory.

Key words: wide baseline images, maximally stable extremal regions on difference of Gaussian space (MSERDoG), affine propagation, adaptive support-weight

摘要： 室外建筑物纹理通常重复而且单一, 在进行宽基线图像匹配时, 得到的初始种子点匹配数量通常比较少, 而且在匹配和扩散时存在匹配多义性问题, 使得应用传统的宽基线准稠密匹配算法不能得到满意的结果。针对这一问题, 提出了一种针对室外建筑物的宽基线图像准稠密匹配算法。算法从高斯差分空间提取最大稳定极值区域, 以获取数量更多的初始种子点匹配; 在仿射传递过程中, 采用自适应支持加权计算匹配分数, 去除匹配多义性问题。实验表明, 提出的算法能获得比较满意的准稠密匹配结果。

关键词: 宽基线图像, 最大稳定极值区域, 仿射传递, 自适应支持加权