Research Seminars

Joint Centre for Digital Music & MMV Group Seminar

Tuesday 9 May 2006, 4:00pm - 6:00pm, Room 105

Talk 1: Structural Segmentation of Musical Audio

Mark Levy, Centre for Digital Music, Queen Mary University of London

Abstract

Segmentation of audio tracks offers improved ways of searching and browsing our increasingly large music collections. We introduce a method of segmenting musical audio into structural sections using a hierarchical model of timbre, and give some experimental results compared to human segmentations. We demonstrate a thumbnailing application and a content-based music search engine based on our segmentation algorithm, and give a brief evaluation of the performance and efficiency of these applications compared to the state of the art.

Talk 2: Multi-feature tracking in a Particle Filter framework

Emilio Maggio, MMV Group, Queen Mary University of London

Abstract

Combining multiple image cues for target modelling has shown promising results in object tracking. However this combination requires a principled method to fuse the cues. In this work the model combines in a single particle filter colour and gradient-based orientation information, together with a highly descriptive representation created by tracking feature points described by the response of rank filters (Ranklets). A reliability measure derived from the particle distribution is used to adaptively weight the contribution of colour and orientation.

Furthermore, information from the tracker is used to set the dimension of the filters for the computation of the gradient, effectively solving the scale selection problem. The position of the feature points is analyzed as well, and a feedback is sent to the global tracker for improved stability. Experiments over a set of real-world sequences show that the adaptive use of colour and orientation information improves over either feature used separately, both in terms of tracking accuracy and of reduction of lost tracks. Also, the automatic scale selection for the derivative filters results in increased robustness.