This page provides information about the large-scale evaluation of the onset detectors discussed in our publication:
Mi Tian, George Fazekas, Dawn Black, Mark Sandler "Design and Evaluation of Onset Detectors Using Different Fusion Policies" to appear in the proceedings of the 15th International Society for Music Information Retrieval Conference (ISMIR 2014), Taipei, Taiwan.
Abstract: Note onset detection is one of the most investigated tasks in Music Information Retrieval (MIR) and various detection methods have been proposed in previous research. The primary aim of this paper is to investigate different fusion policies to combine existing onset detectors, thus achieving better results. Existing algorithms are fused using three strategies, first by combining different algorithms themselves to create new ones, second, by the linear combination of their detection functions, and third, by using a late decision fusion approach. Large scale evaluation is carried out on two published datasets, as well as a new percussion database composed of Chinese traditional instrument samples. An exhaustive search through the parameter space is used enabling a systematic analysis of the impact of each parameter, as well as to report the most generally applicable parameter settings for the onset detectors and the fusion. We demonstrate improved results due to fusion as well as due to the optimised parameter settings.