Coarse-to-fine parameter tuning for content-based object categorization

Object categorization is an interesting application in computer vision. To develop an efficient system for this purpose, finding an appropriate classifier in conjunction with a suitable feature is essential. Most classifiers and features have one or more parameters to be tuned through cross validation. In this paper, we examined a number of classifiers with several feature descriptors and advise an efficient hybrid feature descriptor for object categorization. Besides, we propose a coarse-to-fine parameter tuning method to avoid exhaustive search within various hyper-parameter of the classifiers. The experimental results provided on a subset of COREL dataset shows the efficiency of the advised hybrid feature and the proposed tuning parameters.