Automatic Extraction of Regular Grids from Rectified Facade Image

Automation of facade images interpretation is a crucial task for efficient 3D modeling of urban areas. Information about regularities in facade images can be used for image completion of texture areas occluded with foreground objects, 3D model compression and texture quality enhancement. Existing approaches cannot handle the whole variety of buildings facades and often fail in cases of occluded textures. We present a novel algorithm for automatic extraction of facade structure from single rectified image. We describe facade structure as a set of nonoverlapping grids of similar facade parts, usually containing windows. The most frequent possible grid cells sizes and grid cells centers are estimated by detecting and analyzing similar rectangles in input image. Greedy iterative algorithm sequentially chooses grids with maximum quality measure that incorporates empirical observations about similarities and symmetries. Comparison on the database, created from publically available sources with additional complex examples from our photos, shows that the proposed algorithm outperforms two other state-ofthe-art methods and can handle occlusions.