Correlation-Based Facade Parsing Using Shape Grammar

With strong inference of hierarchical and repetitive structures, semantic information has been widely used in dealing with urban scenes. In this paper, we present a super-pixel-based facade parsing framework which combines the top-down shape grammar splitting with bottom-up information aggregation: machine learning forecasts prior classes, super-pixels improve compactness, and boundary estimation divides the splitting into two procedures - raw and fine, providing a reasonable initial guess for the latter to achieve better random walk optimization results. We also put forward the correlation judging between floors for the purpose of compromising freedom degree reduction with style variety and flexibility, which is also introduced as alignment constraint term to extend the probability energy. Experiments show that our method converges fast and achieves the state-of-the-art results for different styles. Further study on understanding and reconstruction is in progress of exploiting these results.