Modeling Objects with Local Descriptors of Biologically Motivated Selective Attention

Recent work in visual retrieval shows that bag-of-features (BoF) has appeared promising for object recognition and categorization. Local descriptors such as SIFT have shown impressive results on objects. The main idea of BoF is to depict each image as an orderless collection of local keypoint features. However, not all the local keypoint features are important for retrieving objects and rather, the user is often interested in saliency regions of object classes. Therefore, we propose a new method for modeling attention objects with local descriptors. The proposed model in conjunction with a biologically motivated selective attention model can extract the salient regions of each image. In order to model attention objects, we propose a new attention-based SIFT algorithm using scale contract information and local keypoint features to reflect more exact saliency in object classes. Computer experimental results on both Caltech 101 object category datasets and TRECVID2007 datasets shows that the proposed model can generate competitive performance.