Analysis of spatial partitioning approaches for image classification

Abstract

Spatial partitioning is proven to be beneficial for the tasks of image classification, scene categorization and object recognition. The most popular method to capture rough spatial structure of the scene is spatial pyramid matching. However, spatial pyramid matching results in an image representation that is sensitive to rotations. In this research we investigate the influence of upright and rotated partitions on image classification regardless of the image filtering step. We show that simple combination of rotated spatial partitions improves classification accuracy up to 10% compared to single spatial partition commonly used in spatial pyramid matching.