Randomized Selective Search for Locating Object Candidates

Abstract

The effective search for localizing object candidates is a significant method to enhance computational efficiency of object detection and recognition. In this paper, randomized selective search is proposed to improve the hierarchical grouping of neighboring regions and overlap bounding boxes. The main idea is to generate as many potential grouping samples of localizing object candidates as possible with a random neighboring region of the highest similarity. In order to efficiently extract candidates, an output of bounding boxes is selected randomly and combines with its all nearby overlapping boxes. Mean Average Best Overlap (MABO) scores are used to measure the best performance out of all the object candidates. Also, the proposed algorithm is assessed by comparing with the existing method evaluation. Experimental results indicates that the proposed method outperforms the existing one in terms of the quality of object location performance and the quantity of bounding box windows.