Research on Analytic Algorithm of Building Structure Appearance Based on Improved Learning Grammar

Abstract

Semantic segmentation is one of the biggest and most important concerns of computer vision in order to synthesize novel designs and reconstruct buildings. Traditionally, a human expert was required to write grammars for specific building styles, which limited the scope of method applicability. The main purpose of this paper is to improve learning grammar used for building’s façade segmentation. To deal with that, we propose a framework with two layers: in the first layer, we provide a reinforcement learning (RL) techniques to make the segmentation allowing the user to brush strokes on the input image through Gaussian Mixture Models (GMM). Still in this layer, the segmentation can be also make based on shape grammars. Note that for both segmentation, we get as output a ground-truth segmentation. The second layer consist to learn automatically an inferred grammar. Thanks to ground-truth segmentations generated in previous layer, in particular the one generated by RL techniques, we perform clustering techniques to make an improvement of the grammar learned. We evaluate our model on two different datasets and compare in the state-of-the-art our learned-grammar. It show that the proposed outperformed performance gain compared to other learned grammar methods in all the two dataset.