Image representation for object recognition: Utilizing overlapping windows in Spatial Pyramid Matching

Abstract

Image representation holds a very important role in object recognition, as the choice of image representation directly affects the training process and recognition rate. Among many representation technique, Spatial Pyramid Matching (SPM) has been a major breakthrough in the field of object and scene recognition. Expanding from the normal bag-of-words approach, an image is viewed as a pyramid of L + 1 level, where the lth layer of a pyramid is divided into 2l × 2l disjoint sub-windows, and a histogram is extracted from each. In SPM, each image is represented by the concatenation of such histograms. This paper proposes an extension to the normal SPM framework by introducing overlapping sub-windows to replace the disjoint sub-windows of SPM. Two frameworks are proposed in an effort to increase discriminability: rectangular overlapping windows (OW) and circular overlapping windows (CW). We found that the introduction of overlapping sub-windows led to better performances in Caltech 101, Caltech 256 and 15-Scene databases by up to 3.68% under ScSPM and LLC algorithm. Furthermore, it was found that using only the 2nd overlapping layer directly enables us to reduce the memory consumption by 24% while still achieving better recognition rate than traditional SPM.