A Comparative Study of Interactive Segmentation with Different Number of Strokes on Complex Images

Abstract

Interactive image segmentation is the way to extract an object of interest with the guidance of the user. The guidance from the user is an iterative process until the required object of interest had been segmented. Therefore, the input from the user as well as the understanding of the algorithms based on the user input has an essential role in the success of interactive segmentation. The most common user input type in interactive segmentation is using strokes. The different number of strokes are utilized in each different interactive segmentation algorithms. There was no evaluation of the effects on the number of strokes on this interactive segmentation. Therefore, this paper intends to fill this shortcoming. In this study, the input strokes had been categorized into single, double, and multiple strokes. The use of the same number of strokes on the object of interest and background on three interactive segmentation algorithms: i) Nonparametric Higher-order Learning (NHL), ii) Maximal Similarity-based Region Merging (MSRM) and iii) Graph-Based Manifold Ranking (GBMR) are evaluated, focusing on the complex images from Berkeley image dataset. This dataset contains a total of 12,000 test color images and ground truth images. Two types of complex images had been selected for the experiment: image with a background color like the object of interest, and image with the object of interest overlapped with other similar objects. This can be concluded that, generally, more strokes used as input could improve image segmentation accuracy.