A Study on Spatial Information Encoding Scheme for Scene Classification with Local Feature

Recently, spatial pyramid matching (SPM) has achieved leading performance in scene classification. In SPM, spatial information of local features is integrated into representation by computing histograms of local features over increasingly fine regular grids on an image. But this spatial information encoding scheme is very restrictive, because it assumes that images from same scene category have similar layouts, which is always not true. This paper presents a new spatial information encoding scheme for scene classification. Histograms of local features are computed in each bar of horizontal and vertical subdivisions of an image respectively, and the two groups of histograms are concatenated to one overall histogram. This simple spatial information encoding scheme can handle much larger variance in spatial layouts of images than the traditional grid scheme, and also results in a representation with lower dimension. Our approach is compared with SPM on a large dataset of fifteen scene categories. And the experimental results show that our approach can achieve comparable or even higher performance than SPM without combining different fineness of subdivisions.