Image Segmentation Assessment from the Perspective of a Higher Level Task

Image segmentation evaluation is usually performed by visual inspection, by comparing segmentation to a ground-truth, or by computing an objective function value for the segmented image. All these methods require user participation either for manual evaluation, or to define ground-truth, or to embed desired segmentation properties into the objective function. However, evaluating segmentation is a hard task if none of these three methods can be easily employed. Often, higher level tasks such as detecting or classifying objects can be performed much more easily than low level tasks such as delineating the contours of the objects. This fact can be advantageously used to evaluate algorithms for a low level task. We apply this approach to a case study on plankton classification. Segmentation methods are evaluated from the perspective of plankton classification accuracy. This approach not only helps choosing a good segmentation method but also helps detecting points where segmentation is failing. In addition, this more holistic form of segmentation evaluation better meets requirements of big data analysis.