Learning scale ranges for the extraction of regions of interest

Scale space has been widely used in various applications. Given an application, it is essential to decide optimal scales under a certain criterion. Subsampling a scale space is a popular scheme to reduce the search space and thus computational costs. In the context of the extraction of Regions of Interest, we will introduce an alternative scheme that aims to learn scale ranges from training images in order to reduce the search space. We test the proposed scheme in a case study of face localization, and obtain promising results.