Multi-scale Spatial Topic Models for scene recognition
Heping Li · 2016
Some topic models like Latent Dirichlet Allocation (LDA) only use orderless visual-words without considering spatial context information. To model spatial structure for improving scene recognition performance, we propose two novel topic models called Multi-scale Spatial Topic Model (MS-TM) and Multi-scale Spatial Discriminative Topic Model (MS-DTM). Both MS-TM and MS-DTM can effectively fuse multi-scale spatial context information about the visual-words. To estimate the parameters of the proposed models efficiently, we also propose an effective iterative learning method based on variational EM and max-margin idea. Finally, extensive experiments on Li's dataset and LabelMe dataset demonstrate the effectiveness of our proposed models for scene recognition in comparison with other related topic models in the literature.