Dengxin Dai, Hayko Riemenschneider, Gerhard Schmitt, Luc Van Gool
There is an increased interest in the efficient creation of city models, be it virtual or as-built. We present a method for synthesizing complex, photo-realistic facade images, from a single example. After parsing the example image into its semantic components, a tiling for it is generated. Novel tilings can then be created, yielding facade textures with different dimensions or with occluded parts inpainted. A genetic algorithm guides the novel facades as well as inpainted parts to be consistent with the example, both in terms of their overall structure and their detailed textures. Promising results for multiple standard datasets, in particular for the different building styles they contain, demonstrate the potential of the method.
Fig1. The pipeline of our method: From a parsed example facade (a), to its grid representation (b), to a larger, synthesized grid with
inferred label configuration (c), and to the synthesized facade (d). Each node in (b) has a unique label indicating its own tile and it is
highlighted with a specific color. 
Fig2. Comparison of different methods. 
Fig3. An illustration of how the synthesis result evolves with the number of iterations,
resulting in decreasing energy. 
Fig4. Synthesis result on Paris2011 dataset. 
Fig5. Synthesis result on FaSyn13 dataset. 
Fig6. The method is stochastic and is able to generate different solutions from different runs. 
Fig7. Failure cases. The artifacts in (b) are due to the confusing window textures of (a), which results in errors in the semantic labeling
and lattice parsing. In (d), the global symmetric structure of (c) is lost,
which is beyond the ability of our two semi-local constraints.
Fig8. Inpainting Results. Our method is better in keeping structures. 
Fig9. Inpainting Results. The method can come very close to the ground truth. 
This page has been edited by Dengxin Dai