Efficient Reconstruction of Large Unordered Image Datasets for High Accuracy Photogrammetric Applications

The reconstruction of camera orientations and structure from unordered image datasets, also known as Structure and Motion reconstruction, has become an important task for Photogrammetry. Only with few and rough initial information about the lens and the camera, exterior orientations can be derived precisely and automatically using feature extraction and matching. Accurate intrinsic orientations are estimated as well using self-calibration methods. This enables the recording and processing of image datasets for applications with high accuracy requirements. However, current approaches usually yield on the processing of image collections from the Internet for landmark reconstruction. Furthermore, many Structure and Motion methods are not scalable since the complexity is increasing fast for larger numbers of images. Therefore, we present a pipeline for the precise reconstruction of orientations and structure from large unordered image datasets. The output is either directly used to perform dense reconstruction methods or as initial values for further processing in commercial Photogrammetry software.