SIFTing the Relevant from the Irrelevant: Automatically Detecting Objects in Training Images

Many state-of-the-art object recognition systems rely on identifying the location of objects in images, in order to better learn its visual attributes. In this paper, we propose four simple yet powerful hybrid ROI detection methods (combining both local and global features), based on frequently occurring keypoints. We show that our methods demonstrate competitive performance in two different types of datasets, the Caltech101 dataset and the GRAZ-02 dataset, where the pairs of keypoint bounding box method achieved the best accuracies overall.