Spatial pyramid formulation in weakly supervised manner

Abstract

Spatial pyramid match scheme (SPM) is an important scheme in local feature based image classification which effectively adopts geometric structure information into image classification. Most previous approach formulized Spatial Pyramid in unsupervised manner by hierarchical splitting images into separate bins. We found that weak supervised information exists in this process totally unused. We cannot use information directly, because those information corresponding to the combination of all bins, thus we can use those weakly supervised information for bins selection. In this paper, we proposed to select those bins with better discriminative properties.The discriminative property can be well defined from neighborhood entropy. We incorporate local sensitive hash for fast neighborhood identification. We set those bins with higher neighborhood entropy weight zero. Analysis shows that our approach can down weight those non-discriminative bins, in contrast highlighting those discriminative bins. Experiments show that our approach can improve the performance of spatial pyramid match, especially for those categories with complex background. We also proof that under our scheme, result kernel matrix can still preserve positive semi-definite, which can guarantee that our algorithm will coverage. © 2013 Springer-Verlag Berlin Heidelberg.