Weighted pooling of image code with saliency map for object recognition

Recently, codebook-based object recognition methods have achieved the state-of-the-art performances for many public object databases. Based on the codebook-based object recognition method, we propose a novel method which uses the saliency information in the stage of pooling code vectors. By controlling each code response using the saliency value that represents the visual importance of each local area in an image, the proposed method can effectively reduce the adverse influence of low visual saliency regions, such as the background. On the basis of experiments on the public Flower102 database and Caltech object database, we confirm that the proposed method can improve the conventional codebook-based methods.