Example-Based Urban Modeling

Abstract

The manual modeling of virtual cities or suburban regions is an extremely time-consuming task, which expects expert knowledge of different fields. Existing modeling tool-sets have a steep learning curve and may need special education skills to work with them productively. Existing automatic methods rely on rule sets and grammars to generate urban structures; however, their expressiveness is limited by the rule-sets. Expert skills are necessary to typeset rule sets successfully and, in many cases, new rule-sets need to be defined for every new building style or street network style. To enable non-expert users, the possibility to construct urban structures for individual experiments, this work proposes a portfolio of novel example-based synthesis algorithms and applications for the controlled generation of virtual urban environments. The notion example-based denotes here that new virtual urban environments are created by computer programs that re-use existing digitized real-world data serving as templates. The data, i.e., street networks, topography, layouts of building footprints, or even 3D building models, necessary to realize the envisioned task is already publicly available via online services. To enable the reuse of existing urban datasets, novel algorithms need to be developed by encapsulating expert knowledge and thus allow the controlled generation of virtual urban structures from sparse user input. The focus of this work is the automatic generation of three fundamental structures that are common in urban environments: road networks, city block, and individual buildings.<br /> In order to achieve this goal, the thesis proposes a portfolio of algorithms that are briefly summarized next. In a theoretical chapter, we propose a general optimization technique that allows formulating example-based synthesis as a general resource-constrained k-shortest path (RCKSP) problem. From an abstract problem specification and a database of exemplars carrying resource attributes, we construct an intermediate graph and employ a path-search optimization technique. This allows determining either the best or the k-best solutions. The resulting algorithm has a reduced complexity for the single constraint case when compared to other graph search-based techniques. For the generation of road networks, two different techniques are proposed. The first algorithm synthesizes a novel road network from user input, i.e., a desired arterial street skeleton, topography map, and a collection of hierarchical fragments extracted from real-world road networks. The algorithm recursively constructs a novel road network reusing these fragments. Candidate fragments are inserted into the current state of the road network, while shape differences will be compensated by warping. The second algorithm synthesizes road networks using generative adversarial networks (GANs), a recently introduced deep learning technique. A pre- and postprocessing pipeline allows using GANs for the generation of road networks. An in-depth evaluation shows that GANs faithfully learn the road structure present in the example network and that graph measures such as area, aspect ratio, and compactness, are maintained within the virtual road networks. To fill empty city blocks in road networks we propose two novel techniques. The first algorithm re-uses real-world city blocks and synthesizes building footprint layouts into empty city blocks by retrieving viable candidate blocks from a database. We evaluate the algorithm and synthesize a multitude of city block layouts reusing real-world building footprint arrangements from European and US-cities. In addition, we increase the realism of the synthesized layouts by performing example-based placement of 3D building models. This technique is evaluated by placing buildings onto challenging footprint layouts using different example building databases. The second algorithm computes a city block layout, resembling the style of a real-world city block. The original footprint layout is deformed to construct a textit{guidance map}, i.e., the original layout is transferred to a target city block using warping. This guidance map and the original footprints are used by an optimization technique that computes a novel footprint layout along the city block edges. We perform a detailed evaluation and show that using the guidance map allows transferring of the original layout, locally as well as globally, even when the source and target shapes drastically differ. To synthesize individual buildings, we use the general optimization technique described first and formulate the building generation process as a resource-constrained optimization problem. From an input database of annotated building parts, an abstract description of the building shape, and the specification of resource constraints such as length, area, or a number of architectural elements, a novel building is synthesized. We evaluate the technique by synthesizing a multitude of challenging buildings fulfilling several global and local resource constraints. Finally, we show how this technique can even be used to synthesize buildings having the shape of city blocks and might also be used to fill empty city blocks in virtual street networks. All algorithms presented in this work were developed to work with a small amount of user input. In most cases, simple sketches and the definition of constraints are enough to produce plausible results. Manual work is necessary to set up the building part databases and to download example data from mapping services available on the Internet.