Differential edit distance as a countermeasure to video scene ambiguity

In this work the problem of how to evaluate video scene segmentation results is examined. The evaluation, which is typically conducted by comparison of the experimental output of scene segmentation algorithms with a ground-truth temporal decomposition, often suffers from ambiguity in the definition of the ground truth. To alleviate this drawback the use of a string comparison measure, called differential edit distance (DED), is proposed. After defining video scene segmentation evaluation as a string comparison problem, the proposed measure is applied to limit the effect of scene segmentation ambiguity in the performance estimation uncertainty. The experimental results, which include comparisons with state of the art evaluation measures, demonstrate the ambiguity extent and verify the validity of the conducted analysis.