Robust Incremental Structure from Motion

We present a novel method to solve structure and motion problems robustly and incrementally from unordered sets of input images. The proposed method can build large reconstructions without depending on global structure for outlier rejection and starts from the most reliable parts of the data sets. The two main ideas are a strategy to identify reliable subsets of images that have the highest mutual compatibility and an ordering of the reconstruction buildup that gives higher priority to these subsets and merges new information according to this ordering. Another advantage of our buildup strategy is that loop closing is done unbiased by drift in spite of adding correspondence data incrementally. Correspondence information is only locally verified and merged into reconstructions without immediately removing outliers based on globally reconstructed structure. We demonstrate the robustness and scalability of our approach on several large reconstructions from unordered sets of images and indicate the achieved accuracy by preserving the topology of the 3D structure and cameras.