Visual analysis for drum sequence transcription

Abstract

A system is presented for analysing drum performance video sequences. A novel ellipse detection algorithm is introduced that automatically locates drum tops. This algorithm fits ellipses to edge clusters, and ranks them according to various fitness criteria. A background/foreground segmentation method is then used to extract the silhouette of the drummer and drum sticks. Coupled with a motion intensity feature, this allows for the detection of ‘hits’ in each of the extracted regions. In order to obtain a transcription of the performance, each of these regions is automatically labeled with the corresponding instrument class. A partial audio transcription and color cues are used to measure the compatibility between a region and its label, the Kuhn-Munkres algorithm is then employed to find the optimal labeling. Experimental results demonstrate the ability of visual analysis to enhance the performance of an audio drum transcription system.