CREATING PROCEDURAL WINDOWBUILDING BLOCKS USING THE GENERATIVE FACT LABELING METHOD

CREATING PROCEDURAL WINDOW BUILDING BLOCKS USING THE GENERATIVE FACT LABELING METHOD W. Thallera , R. Zmugga , U. Krispela , M. Poscha , S. Havemanna , D.W. Fellner a,b a Institute of ComputerGraphics & KnowledgeVisualization (CGV), TU Graz, Austria b TU Darmstadt & Fraunhofer IGD, Germany KEY WORDS: Procedural Modeling, Neo-Classical Architecture, 3D-Reconstruction, Shape Grammars, Fact Labeling Method ABSTRACT: The generative surface reconstruction problem can be stated like this: Given a finite collection of 3D shapes, create a small set of functions that can be combined to generate the given shapes procedurally. We propose generative fact labeling (GFL) as an attempt to organize the iterative process of shape analysis and shape synthesis in a systematic way. We present our results for the reconstruction of complex windows of neo-classical buildings in Graz, followed by a critical discussion of the limitations of the approach. • Synthesis phase: We have produced a library of combineable procedural assets corresponding to the elements identified in the analysis phase (see Section 4). • Verification phase: We have 3D-reconstructed several wellchosen windows from this collection in order to assess the usefulness of our procedural library (see Section 5). We discuss strengths, weaknesses and limitations in Section 6. 1.1 Figure 1: Reasons for the complexity of window modeling. Intricacies of facade composition (left), vertical coherence (middle), and horizontal coherence (right) between ajacent windows. 1 Contribution We introduce the mentioned concepts (fact label, attribute, element, procedural asset, exemplar) as part of our generative fact labeling (GFL) method, a simple conceptual framework to deal with families of complex structured shapes. The goal of this method is generative shape reconstruction, i.e., to produce a library of functions that allow not only reproducing the limited number of given exemplars, but also the design space that is spanned by them. Whenever factoring shapes into procedural elements or components, one must be aware that this factorization is only an interpretation (speculation); there is no such thing as the “best” procedural description. Elegance is related to simplicity, but one can never be sure, e.g., to have found the shortest procedural description1 . Our method is a guideline how to find at least a reasonable procedural explanation of a complex shape class. INTRODUCTION Most digital urban reconstructions today suffer from bad windows. There are two main sources of inaccuracy: Either the window is well modeled but does not match the original (because it is selected from a set of pre-modeled assets), or the window matches but is badly modeled (window texture). How can this situation be improved? We opt for geometric reasoning. Windows are among the most salient features of façades. In most classical styles of architecture a window is not just a rectangular hole in a wall, but rather a combination of different inter-related design elements. They may derive from a long-standing architectural tradition. Thus, when creating a 3D model of a façade, a substantial part of the effort will be spent on modeling the windows and the decorative elements that go with them. A library of common pre-modeled windows can be used only for superficial reconstructions. When more accuracy is required, windows from an asset library can at most be used as a starting point for further manual modeling, or they must be camouflaged by photo-texturing. Windows in different buildings are often similar but hardly ever identical. We must better understand this phenomenon (Figures 1, 2) to produce more accurate reconstructions. 1.2 Benefit Procedural models have striking advantages over other types of 3D models (compact, editable, re-usable, scalable, meaningful semantics), especially over 3D scans (sampling approach). Generating 3D models by varying a few high-level parameters is nice, but it can be difficult to determine the parameters of a given realworld shape (procedural shape fitting). Even more challenging, and far from solved, is the aforementioned problem to determine a suitable set of procedures for a given set of shape examplars (also called inverse procedural modeling). The use case in this paper shows how to approach such a complex problem. In this paper we propose a methodological approach to deal with situations where a large number of highly structured, similar but not identical shapes must be captured. Our generative fact labeling (GFL) method has three phases: The second benefit is that we invite all knowledgeable specialists in the field of architecture to refine and specialialize our imperfect taxonomy (discussion in Section 6). In contrast to other taxonomies with mainly academic value, the purpose of ours is to actually reproduce the shapes, i.e., it is a generative taxonomy. • Analysis phase: We have gathered a collection of 150 photographs of complex windows. We have structured them into elements by assigning fact labels (see Section 3). 1 The Kolmogorov complexity KC(b) of a bit sequence b, the length of the shortest computer program that (re-)produces b, is not computable. 235 Figure 2: Window exemplars. The full set contains 150 images of windows from neo-classical buildings erected in Graz, Austria, in 1860-1890 (Gründerzeit). The complexity and visual dominance of the windows pose challenges to any digital urban reconstruction. 2 RELATED WORK 3 WINDOW ANALYSIS Digital reconstruction of buildings and monuments was a prominent topic throughout the years, and generative approaches on architecture are a hype today. Already the first shape grammars (Stiny and Gips, 1972, Stiny, 1980) were very useful for understanding the patterns of classical architecture. Hierarchical structures, such as façade layouts, are an ideal use case for shape grammars; but it turns out that their value for windows is only limited. In the beginning we are confronted with an unordered set of about 150 exemplars of complex windows, a selection of which is depicted in Figure 2. The question now is, how do we synthesize a function library that reproduces these windows? As outlined before, the first step is the analysis described in the following. One of the first applications of shape grammars not just to understand, but also to generate complex architecture was (Wonka et al., 2003). Recent approaches on shape grammars for procedural modeling of architecture (Müller et al., 2006, Hohmann et al., 2010, Krecklau et al., 2010) usually focus on scripting as the main method to achieve their stunning results. As an alternative we have presented in (Zmugg et al., 2012) an approach where scripted procedural assets and shape functions can be applied and assembled interactively in order to reconstruct complicated architecture (e.g. the façade of the Louvre in Paris). The system presented in this paper follows the same approach. The result of the mode (...truncated)