Discovering repetitive patterns in facade images using a RANSAC-style algorithm

In this paper, we present an algorithm that automatically decomposes the facade images of buildings into floors and tiles by discovering the repetitive patterns of the dominant structures, such as windows and balconies. Our algorithm follows a histogram-based approach that analyzes the accumulated horizontal and vertical histogram profiles of window gradients and edges. In this study, a histogram is viewed as a series of noisy wave cycles, where a wave cycle represents the approximated position and dimensions of a window. Therefore, the dominant frequency of the histogram should be highly related to the windows. The repetitive pattern that represents these windows is then discovered from the dominant frequency by iteratively fitting candidate sine waves using a RANSAC-styled algorithm. Finally, the splitting lines are positioned in the valleys of the resultant wave. We evaluate our algorithm using the publicly available facade image database, and the results demonstrate that our proposed algorithm performs well. Comparisons between the proposed algorithm and several baseline techniques are also evaluated and discussed.