Still Image Objective Segmentation Evaluation using Ground Truth

In this paper, an objective segmentation evaluation metric suitable for the evaluation of still image segmentation results is proposed. The proposed metric is based on the spatial accuracy approach, originally proposed for the evaluation of foreground/backgroung segmentation masks generated from video sequences. This approach is extended to still image segmentation evaluation, where both the estimated segmentation masks and the ground truth mask typically contain multiple regions. The proposed method takes into account, using a single metric, not only the accuracy of the boundary localization of the created segments but also the under-segmentation and over-segmentation effects, which can hinder the performance of any segmentation algorithm and decrease the usability of the segmentation results in content-based applications. Several experiments have shown the potential of this approach.