Bag-of-Words and Region-Based Feature Representations in Object Categorization: A Comparative Study

The aim of object categorization is to find a given object in an image and the performance of object categorization heavily depends on the extracted features as the image descriptor. In the literature, feature representation can be broadly classified into block/region-based and bag-of-words (BoW) features. However, there is no a comparative study of using these different feature representations over different datasets and different image scales since the image sizes for object recognition are varying from different datasets. Our experimental results using the Corel and PASCAL datasets show that when images contain more complex scenes like Corel images, the block-based feature is a better choice. In addition, the larger the image scales, the better the recognition performance. On the contrary, when images contain fewer objects like PASCAL images, it is better to consider the region-based feature representation. Particularly, reducing the image scale does not degrade the recognition performance; it even shows some level of improvement. On the other hand, although the BoW feature does not perform better than the block/region based features, it shows stable performances over different datasets and different image scales. This indicates that when the chosen dataset contains a large amount of images having various types of contents, which is difficult to decide what features to be extracted, the BoW feature can be extracted as the baseline feature representation.