Utilizing Overlapping Windows in Spatial Pyramid Matching

Spatial Pyramid Matching (SPM) has been a major breakthrough in the field of object and scene recognition. Using this approach, an image is divided into 2 × 2 disjoint sub-windows for each pyramid level l. However, the disjoint arrangement of the sub-windows can be too restrictive, especially when we consider that each window may benefit from broader context. Inspired by human habit in viewing an image, we introduced two overlapping spatial windows in this paper: rectangular overlapping windows (OWSPM) and circular overlapping windows (CWSPM). We found that the introduction of overlapping spatial window allows us to achieve better performances in Caltech 101, Caltech 256 and 15-Scene databases (up to 3.68% compared to traditional SPM using ScSPM algorithm). Furthermore, it enables us to bypass the 0 and 1 layer and use the 2 pyramid layer directly for recognition, considerably cutting memory consumption while achieving better recognition rate than traditional SPM at the same time.