ABSTRACT
Kernel-based methods have proved highly effective in many applications because of their wide generality. This chapter proposes a general framework to build positive-definite kernels between images, allowing us to leverage, e.g., shape and appearance information to build meaningful similarity measures between images. It introduces positive-definite kernels between appearances as coded in region adjacency graphs, with applications to classification of images of object categories. The chapter also introduces positive-definite kernels between shapes as coded in point clouds, with applications to classification of line drawings. It proposes to model appearance of images using region adjacency graphs obtained by morphological segmentation. It compares test error rate performances of SVM-based multi-class classification with histogram kernel (H), walk kernel (W), tree-walk kernel (TW), and the tree-walk kernel with weighted segments (wTW). The chapter suggests that histograms do not carry any supplemental information over walk-based kernels: the global histogram information is implicitly retrieved in the summation process of walk-based kernels.
