Saliency weighted Spatial Pyramid Representation for object recognition

In this paper, we propose a method based on boolean-map saliency to improve the spatial pyramid pooling technique for object recognition. Spatial Pyramid Representation with bag-of-words model has been remarkably successful in terms of generic image recognition, which has become a standard feature pooling step in image recognition procedure employed by many state-of-the-art methods. On the other hand, visual saliency method based on the Gestalt psychological studies and the Boolean Map theory has the advantage of perceiving structural information in image scene without any prior knowledge. Our work focus on the application of Boolean Map model in object recognition, using an attention map to weight the spatial pyramid for a better representation. The method is evaluated on three datasets to show the performance improvement comparing with the original model.