Image classification using latent spatial pyramid matching

We present work on image classification in this thesis. Image classification is a classical task in computer vision, whose goal is to determine whether or not any instances of a particular object class appear in a given image. There are three major pieces of work. First, we proposed a novel Latent Spatial Pyramid Matching (L-SPM) feature representation inspired by the state-of-art Spatial Pyramid Matching (SPM) [29] feature representation. L-SPM allows the cells of the pyramid to move within reasonable regions instead of a predefined rigid partition. Second, we utilize Efficient Subwindow Search [28] based on a branch-and-bound algorithm to select the position and size for the latent cells. Third, we implement the Latent SVM framework proposed by Felzenszwalb et al. [21] to solve the non-convex optimization problem. Results are reported for image classification on the Pascal VOC 2007 data set.