Unsupervised cosegmentation based on superpixel matching and Fastgrabcut
Hongkai Yu, Xiaojun Qi · 2014
This paper proposes a novel unsupervised cosegmentation method which automatically segments the common objects in multiple images. It designs a simple superpixel matching algorithm to explore the inter-image similarity. It then constructs the object mask for each image using the matched superpixels. This object mask is a convex hull potentially containing the common objects and some backgrounds. Finally, it applies a new FastGrabCut algorithm, an improved GrabCut algorithm, on the object mask to simultaneously improve the segmentation efficiency and maintain the segmentation accuracy. This FastGrabcut algorithm introduces preliminary classification to accelerate convergence. It uses Expectation Maximization (EM) algorithm to estimate optimal Gaussian Mixture Model(GMM) parameters of the object and background and then applies Graph Cuts to minimize the energy function for each image. Experimental results on the iCoseg dataset demonstrate the accuracy and robustness of our cosegmentation method.