Keybook: Unbias object recognition using keywords. Issue 8 (15th May 2015)

• Record Type: Journal Article
• Authors: Hoo, Wai Lam; Lim, Chern Hong; Chan, Chee Seng
• Is Part Of: Expert systems with applications. Volume 42:Issue 8(2015)
• Page Start: 3991
• Page End: 3999
• Publication Date: 2015-05-15
• Subjects: Dataset bias -- Object recognition -- Codebook generation -- Bag-of-Words model; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2015.01.019
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Current object recognition systems suffer from dataset bias issue. Keybook is proposed to undo the dataset bias problem. Mutual information is employed to assist the generation of Keybook. Empirically, the proposed method shows improvements over conventional solutions. Future work include extension to human motion analysis and fine-grain classification. Abstract: The presence of bias in existing object recognition datasets is now a well-known problem in the computer vision community. In this paper, we proposed an improved codebook representation in the Bag-of-Words (BoW) approach by generating Keybook. In specific, our Keybook is composed from the keywords that significantly represent the object classes. It is extracted utilizing the concept of mutual information. The intuition is to perform feature selection by maximize the mutual information of the features between the object classes; while minimize the mutual information of the features between the domains. With this, the Keybook will not bias to any of the domain and consists of valuable keywords among the object classes. The proposed method is tested on four public datasets to evaluate the classification performance in seen and unseen datasets. Experiment results have showed the effectiveness of our proposed methods in undo the dataset bias problem.