Abstract: One of the most efficient ways to tackle the scalability problem in large scene reconstruction is to break apart the scene into a number of sub-problems, then reconstruct each sub-problem independently, and merge the partial reconstructions finally. Image clustering without any camera or scene prior information is a difficult problem in 3D reconstruction. Image clustering is inherently time consuming, and generally no satisfactory results can be achieved. This paper explores how to use the auxiliary information of cameras which is non-accurate and neglected but available, and substantially simplifies the image clustering in 3D scene reconstruction. Firstly, the view-overlap is computed, then the view-overlap based clustering approach is proposed, and finally the clusters are independently reconstructed and merged. The experiments are done on several sets of real images, the results show that compared with the image retrieval based method and Samantha method, the scene reconstruction via camera auxiliary information performs satisfactorily in terms of efficiency, robustness and scalability.

Key words: auxiliary information, graph partitioning, view clustering, 3D reconstruction

摘要： 将图像先分组重建，然后融合的方法，是解决大场景三维重建中规模问题的最有效的途径。无任何先验信息下的图像分组，不仅计算量大，而且很难取得有效的分组结果。对如何利用相机中一些精度不高甚至被人们所忽略的粗略辅助信息问题进行了研究，简化了大场景三维重建中快速、鲁棒的分组重建问题。首先借助辅助信息进行视图间的重叠度计算，并据此进行了视图间的分组，最后完成了视图组内重建和组间融合。在几组真实的图像上进行了实验测试，结果表明，借助辅助信息的分组重建方法相比基于图像检索的方法和Samantha方法，在效率和重建的鲁棒性上都有一定的优势。

关键词: 辅助信息, 图划分, 视图聚类, 三维重建