Enhanced multi-view dancing videos synchronisation

This paper describes a system for automatically synchronising multi-view video sequences of Salsa dancing recorded with multimodal capturing platform. The multimodal capturing setup consists of audiovisual streams along with depth maps and inertial measurements. Part of the dataset was video sequences captured from machine vision cameras and Microsoft Kinect sensor that were not temporal synchronised during the capturing stage. As an essential step, we proposed efficient solutions for synchronisation of these data based on co-occurrence appearance changes. In order to improve the accuracy, the proposed system employed state-of-art body detection and tracking algorithm to obtain Region of Interest, within which the appearance changes are analysed. The accurately synchronised video set can then be further analysed and augmented for visualisation and evaluation of dancing performance.