Bottom-up Estimation of Geometric Layout for Indoor Images

In this paper, we propose a bottom-up approach to estimate the geometric layout of indoor images using latent variables. By utilizing latent variables to model subregions, the estimation accuracy of scene layout is implicitly improved. The proposed method consists of three sub-tasks: feature extraction, subregion classification and geometric layout classification. Firstly, the location features are extracted to roughly estimate the basic indoor structure. The influence of illumination, rich color, and foreground occlusion can be eliminated. Secondly, N-slack SSVM is applied to efficiently classify the location features extracted in the previous step. Finally, the bag-of-words model is combined with cosine similarity and information divergence filtering to improve the fault tolerance of the geometric layout classification task. The classification accuracy can reach 0.982, which well demonstrate the effectiveness of the proposed approach.