Region Proposal Generation: A Hierarchical Merging Similarity-Based Algorithm
Paper ID : 1145-MVIP2020
Maryam Taghizadeh *, Abdolah Chalechale
This paper presents a hierarchical algorithm usingregion merging with the aim of achieving a powerful pool ofregions for solving computer vision problems. An image is firstrepresented by a graph where each node in the graph is asuperpixel. A variety of features is extracted of each region, whichis next merged to neighbor regions according to the new algorithm.The proposed algorithm combines adjacent regions basedon a similarity metric and a threshold parameter. By applyingdifferent amounts for the threshold, a wide range of regions isacquired. The algorithm successfully provides accurate regionswhile they can be represented through the bounding box andsegmented candidates. To extensively evaluate, the effectivenessof features and the combination of them are analyzed on MSRCand VOC2012 Segmentation dataset. The achieved results areshown a great improvement at overlapping in comparison tosegmentation algorithms. Also, it outperforms previous regionproposal algorithms, especially it leads to a relatively great recallat higher overlaps (>0.6).
merging, region proposal, superpixel, segmentation
Status : Paper Accepted (Oral Presentation)