A New Data Transformation Method Based on Adaptive Binarization for Bag-of-Features Model

Visual object categorization has gained more and more attention in computer vision and bag-of-features model has become an important approach to form an object categorization system. As for image feature representation, "continuous valued" histogram that records frequency of each visual word and "binarized value" histogram that records only absence/presence of each visual word are commonly used in bag-of-features model. In this paper, we utilize the merits of both representations and propose a scheme to use adaptive binarization method to transform "continuous valued" histogram to "binarized value" histogram. Experiments on "The PASCAL Visual Object Classes Challenge 2006" show that this data transformation approach improves the performance of object categorization system.