High-Level Facade Image Interpretation using Marked Point Processes

In this thesis, we address facade image interpretation as one essential ingredient for the generation of high-detailed, semantic meaningful, three-dimensional city-models. Given a single rectified facade image, we detect relevant facade objects such as windows, entrances, and balconies, which yield a description of the image in terms of accurate position and size of these objects. Urban digital three-dimensional reconstruction and documentation is an active area of research with several potential applications, e.g., in the area of digital mapping for navigation, urban planing, emergency management, disaster control or the entertainment industry. A detailed building model which is not just a geometric object enriched with texture, allows for semantic requests as the number of floors or the location of balconies and entrances. Facade image interpretation is one essential step in order to yield such models. In this thesis, we propose the interpretation of facade images by combining evidence for the occurrence of individual object classes which we derive from data, and prior knowledge which guides the image interpretation in its entirety. We present a three-step procedure which generates features that are suited to describe relevant objects, learns a representation that is suited for object detection, and that enables the image interpretation using the results of object detection while incorporating prior knowledge about typical configurations of facade objects, which we learn from training data. According to these three sub-tasks, our major achievements are: We propose a novel method for facade image interpretation based on a marked point process. Therefor, we develop a model for the description of typical configurations of facade objects and propose an image interpretation system which combines evidence derived from data and prior knowledge about typical configurations of facade objects. In order to generate evidence from data, we propose a feature type which we call shapelets. They are scale invariant and provide large distinctiveness for facade objects. Segments of lines, arcs, and ellipses serve as basic features for the generation of shapelets. Therefor, we propose a novel line simplification approach which approximates given pixel-chains by a sequence of lines, circular, and elliptical arcs. Among others, it is based on an adaption to Douglas-Peucker’s algorithm, which is based on circles as basic geometric elements. We evaluate each step separately. We show the effects of polyline segmentation and simplification on several images with comparable good or even better results, referring to a state-of-the-art algorithm, which proves their large distinctiveness for facade objects. Using shapelets we provide a reasonable classification performance on a challenging dataset, including intra-class variations, clutter, and scale changes. Finally, we show promising results for the facade interpretation system on several datasets and provide a qualitative evaluation which demonstrates the capability of complete and accurate detection of facade objects.