Skip to Main Content
In this paper, we explore the segmentation performance of four state-of-the-art unsupervised over-segmentation methods by evaluating the quality of over-segments. To achieve this goal, we design a novel real-time evaluating framework, this method is integrating the classical goodness methods and our proposed new discrepancy method based on real-time boundary extraction, which is able to suppress the minor edges found within a homogeneous region, while being able to locate the important edges in real-time. We use it to qualitatively and quantitatively evaluate the quality of over-segments. Experiments are carried out both on the Berkeley Segmentation Dataset and real-scene images. The quality-evaluating results show that the EGIS outperforms the other three over-segmentation methods. The evaluating accuracy and efficiency of our proposed method is verified by the over-segmentation results.