Improvements to the Descriptor of SIFT by BOF Approaches

The efficacy and efficiency of SIFT have made it a state-of-art feature descriptor. It has been widely used in many computer vision applications such as image classification. A large number of methods, e.g. PCA-SIFT, have been contributed to further improve its performance focusing on different components of it. Differing from those previous works, we broach a new scheme to improve the performance of SIFT's descriptor in this paper. We first establish the connection between SIFT and bag of features (BOF) model in descriptor construction. Based on this connection, we then introduce approaches of BOF, e.g. the preservation of spatial information (we adopt spatial pyramid matching as an example to achieve this goal), into SIFT to enhance its robustness. Experimental results in scene matching and image classification show that the BOF-driven SIFT effectively and consistently outperforms the original SIFT.