Region Proposal Generation: A Hierarchical Merging Similarity-Based Algorithm

This paper presents a hierarchical algorithm using region merging with the aim of achieving a powerful pool of regions for solving computer vision problems. An image is first represented by a graph where each node in the graph is a superpixel. A variety of features is extracted of each region, which is next merged to neighbor regions according to the new algorithm. The proposed algorithm combines adjacent regions based on a similarity metric and a threshold parameter. By applying different amounts for the threshold, a wide range of regions is acquired. The algorithm successfully provides accurate regions while can be represented through the bounding box and segmented candidates. To extensively evaluate, the effectiveness of features and the combination of them are analyzed on MSRC and VOC2012 Segmentation dataset. The achieved results are shown a great improvement at overlapping in comparison to segmentation algorithms. Also, it outperforms previous region proposal algorithms, especially it leads to a relatively great recall at higher overlaps (≥ 0.6).