The quality of the parameters of relative orientation (ROs) of a stereoscopic pair of images is crucial for the quality of results in structure from motion (SfM). In this paper we focus on improving the robustness and accuracy of SfM by detecting and eliminating incorrect ROs, especially due to repetitive structure, that typically result in incorrectly estimated RO results and thus degrade 3D reconstruction. ROs are represented by a view graph. We develop a novel variant of cycle consistency inference, called sequential cycle consistency inference or SCCI, to infer wrong edges by analyzing the geometric consistency of cycles in the view graph. Our method consists essentially of a two-stage process, an initialization step based on the union of various orthogonal minimum spanning trees (MST) of the graph, followed by an expansion step which incrementally adds edges to this graph. We show by means of experimental studies that our method reaches better robustness and accuracy for data sets containing repetitive structure, compared to state-of-the-art algorithms of RO outlier detection.

立体图像对的相对方向 (RO) 参数的质量对于运动结构 (SfM) 结果的质量至关重要。在本文中，我们专注于通过检测和消除不正确的 RO 来提高 SfM 的鲁棒性和准确性，尤其是由于重复结构，这通常会导致错误估计的 RO 结果，从而降低 3D 重建。RO 由视图图表示。我们开发了一种循环一致性推断的新变体，称为顺序循环一致性推断或 SCCI，通过分析视图图中循环的几何一致性来推断错误的边缘。我们的方法本质上由一个两阶段过程组成，一个基于图的各种正交最小生成树（MST）联合的初始化步骤，然后是一个扩展步骤，该步骤逐渐向该图添加边。我们通过实验研究表明，与 RO 异常值检测的最新算法相比，我们的方法对包含重复结构的数据集具有更好的鲁棒性和准确性。