Region pyramid based hierarchical region selection for efficient object detection

Abstract

In this paper we propose a novel hierarchical region selection method for object detection based on region pyramid. The significant contribution of this paper is the introduction of a general purpose candidate region selector that accelerates object detectors. To reduce the search space, a region pyramid that represents the hierarchical relationship between candidate regions is proposed. By jumping and pruning the region pyramid based on quality bound and detection results, candidate regions that have no possibility of being an object can be skipped over efficiently. A set of experiments on various dataset is presented. The proposed method greatly reduces the overall computation time and reduces the number of false positives.