Are you an EPFL student looking for a semester project?
Work with us on data science and visualisation projects, and deploy your project as an app on top of GraphSearch.
Terrestrial imagery found in public databases is an alternative and complementary data source for environmental studies. Compared to usual data source, such as airborne images, terrestrial pictures may have higher temporal and spatial resolutions. They are especially valuable in studying the past, when airborne acquisitions were more rare or inexistent. Nevertheless, terrestrial image databases are typically imperfectly georeferenced, if at all. Hence, it is difficult to retrieve images of interest in a database. Moreover, georeferencing processes proposed for these images are either totally manual or rely on the existence of a collection of similar and overlapping images. The first is time-consuming and the second is appropriate only for particular sets of images. An alternative and more general georeferencing procedure is the registration of a picture with a landscape model. In this thesis, we propose to use as the reference a 3D virtual globe and our aim is to retrieve the orientation and the location of the camera. The location and orientation of a picture can be computed with correspondences that a user recognizes and clicks in a picture and in an orthoimage. However, these two distinct types of images are difficult to match because of the distortions resulting from the different viewpoints. Hence, we provide a navigation tool to scroll through a 3D model to ease this task. Indeed, to find the orientation of a picture, users typically exploit the skyline which appears only in a 3D view. In imitation, we develop a novel skyline-matching technique, based on Dynamic Time Warping, to compute the fine orientation of a picture. In a second phase, we assess whether the similarities between a 3D synthetic image and real images can be used to compute the orientation of a picture. The matching of real images with a model is challenging. We show how images shared on the web can be inserted in the landscape model to ease the georeferencing. Finally, all of the presented pose estimation schemes assume that every location of the map is as likely to be a shooting spot. However, we show that geographic indicators related to accessibility and morphology are correlated with a database of picture locations. Exploiting these relations, our last contribution is a map of the attractiveness. In this thesis, we observed that pose estimation from the automatic registration of a real image with a landscape model cannot compete with the accuracy reached by an operator or a standard photogrammetric workflow. However, the proposed methods can greatly reduce the operator workload for the manual georeferencing by finding automatically correspondences with a 3D landscape model. Second, they are a valid alternative if the required conditions for Structure-from-Motion are not respected. Finally, the accuracy obtained is sufficient to explore image databases with spatial queries and to add geographic information in the pictures. Registration of landscape images with landscape models has previously received limited attention. The tools developed in this thesis can ease the work of archivists (in the management and provision of efficient image databases), environmental scientists (for the extraction of georeferenced information from pictures) and the general public (to browse collections of augmented images inserted in a virtual globe).
Frédéric Kaplan, Isabella Di Lenardo, Rémi Guillaume Petitpierre, Beatrice Vaienti
Maud Ehrmann, Frédéric Clavert