Analysis of Automatic Multi-label GrabCut using NPR for natural image segmentation

Abstract

Automatic Multi-label GrabCut is an extension of the standard GrabCut technique to segment a given image automatically into its natural segments without any user intervention. The Normalized Probabilistic Rand (NPR) index is able to give meaningful comparisons by comparing different images and different segmentations of the same image. In this paper, more analysis is conducted to evaluate the efficiency of the developed automatic multi-label GrabCut using the NPR index. Based on using more than one human ground truth, segmentations are conducted on a large scale of the Berkeley's benchmark of natural images. The NPR, PR and GCE metrics produced acceptable accuracy measures emphasizing the scalability of the proposed technique for large scale datasets. Comparisons are applied for different images and experiments show that the NPR is the most efficient score to determine good segmentation compared to other metrics.