Image Segmentation by Cascaded Superpixel Merging with Privileged Information

Abstract

We propose a learning-based image segmentation algorithm. Starting from superpixels, our method learns the probability of merging two regions based on the ground truth made by humans. The learned information is used in determining whether the two regions should be merged or not in a segmentation stage. Unlike exiting learning-based algorithms, we use both local and object information. The local information represents features computed from super pixels and the object information represent high level information available only in the learning process. The object information is considered as privileged information, and we can use a framework that utilize the privileged information such as SVM+ and AdaBoost+. In experiments on the Berkeley Segmentation Dataset and Benchmark( BSDS500) and PASCAL Visual Object Classes Challenge(VOC 2012) data set, our model exhibited the best performance with a relatively small training data set and also showed competitive results with a su ciently large training data set.