Hierarchical Evaluation of Segment Boundary Detection

Structure in music is traditionally analyzed hierarchically: large-scale sections can be sub-divided and refined down to the short melodic ideas at the motivic level. However, typical algorithmic approaches to structural annotation produce flat temporal partitions of a track, which are commonly evaluated against a similarly flat, human-produced annotation. Evaluating structure analysis as represented by flat annotations effectively discards all notions of structural depth in the evaluation. Although collections of hierarchical structure annotations have been recently published, no techniques yet exist to measure an algorithm’s accuracy against these rich structural annotations. In this work, we propose a method to evaluate structural boundary detection with hierarchical annotations. The proposed method transforms boundary detection into a ranking problem, and facilitates the comparison of both flat and hierarchical annotations. We demonstrate the behavior of the proposed method with various synthetic and real examples drawn from the SALAMI dataset.