A Ground-Truth Fusion Method for Image Segmentation Evaluation

Image segmentation evaluation is popularly categorized into two different approaches based on whether the evaluation uses a human expert's manual segmentation as a reference or not. When comparing automated segmentation against manual segmentation, also referred to as the ground-truth segmentation, multiple ground-truths are usually available. Much research has been done on analysis of segmentation algorithms and performance metrics, but very little study has been done on analyzing techniques for ground-truth fusion from multiple ground-truth segmentations. We propose a hybrid ground-truth fusion technique for image segmentation evaluation and compare it with other existing ground-truth fusion methods on a data set having multiple ground-truths at various coarseness levels. Qualitative and quantitative results show that the proposed method provides improved segmentation evaluation performance.