Combining complementary kernels in complex visual categorization

Bag Of Words model [1] and Fisher Vectors [2] coupled with incorporated spatial information, such as the Spatial Pyramid Matching (SPM) [3] or the spatial Fisher vectors [4], proved to reach state of the art performances in many Image categorization tasks, e.g. the PASCAL VOC challenge [5]. However, image categorization remains a very challenging task because most descriptors present strong intra-class variabilities and inter-class correlations