Promotie A. Bazzica: muziek
15 december 2017 12:30 - Locatie: Aula, TU Delft - Door: webredactie
Music information retrieval beyond audio. A vision-based approach for real-world data. Promotor: Prof.dr. A. Hanjalic (EWI).
Music is not just experienced by listening to audio recordings. We also watch it, for instance at a live concert or on YouTube. From a multimedia content analysis perspective, at first video recordings of a music performance look messy and hard to be exploited, because they consist of unconstrained and unstructured visual content. However, if we find ways to understand complex scenes and extract musically relevant cues for each featuring musician, the existing digital music platforms can be further innovated bringing a more informative and non-linearly accessible experience to users which accommodates multiple perspectives on the content.
This thesis sheds light on the challenges of real-world data analysis for vision-based Music Information Retrieval; this is done by presenting a feature named Motion Orientation Histograms for unsupervised musician movements analysis, a score reduction method based on Playing/Non-Playing matrices to be used for performance-to-score synchronization, a semi-automatic musician-wise Playing/Non-Playing annotation system, a face clustering method for quasi-static scene videos, and a supervised approach based on multiple-stream convolutional networks for note onsets detection of clarinetist videos.
Even though the vision-based Music IR field remains a largely unexplored territory, which requires multi-disciplinary research, our outcomes show the benefit of tackling the challenges of real-world data. As opposed to oversimplified data, unstructured and unconstrained music videos reveal concrete opportunities to deploy new applications on digital music platforms.
Voor inzage in proefschriften van de promovendi kunt u kijken in de TU Delft Repository, de digitale vindplaats van openbare publicaties van de TU Delft. Proefschriften zullen binnen een paar weken na de desbetreffende promotie in de Repository te vinden zijn.