Image segmentation and its performance evaluation are very difficult but important problems in computer vision. A major challenge in segmentation evaluation comes from the fundamental conflict between generality and objectivity: For general-purpose segmentation, the ground truth and segmentation accuracy may not be well defined, while embedding the evaluation in a specific application, the evaluation results may not be extensible to other applications. We present a new benchmark to evaluate five different image segmentation methods according to their capability to separate a perceptually salient structure from the background with a relatively small number of segments. This way, we not only find a large variety of images that satisfy the requirement of good generality, but also construct ground-truth segmentations to achieve good objectivity. We also present a special strategy to address two important issues underlying this benchmark: (1) most image-segmentation methods are not developed to directly extract a single salient structure; (2) many real images have multiple salient structures. We apply this benchmark to evaluate and compare the performance of several state-of-the-art image segmentation methods, including the normalized-cut method, the watershed method, the efficient graph-based method, the mean-shift method, and the ratio-cut method.