Toward automated evaluation of interactive segmentation

We previously described a system for evaluating interactive segmentation by means of user experiments (McGuinness and O’Connor, 2010). This method, while effective, is time-consuming and labor-intensive. This paper aims to make evaluation more practicable by investigating if it is feasible to automate user interactions. To this end, we propose a general algorithm for driving the segmentation that uses the ground truth and current segmentation error to automatically simulate user interactions. We investigate four strategies for selecting which pixels will form the next interaction. The first of these is a simple, deterministic strategy; the remaining three strategies are probabilistic, and focus on more realistically approximating a real user. We evaluate four interactive segmentation algorithms using these strategies, and compare the results with our previous user experiment-based evaluation. The results show that automated evaluation is both feasible and useful