Video segmentation using BIC and stacked scanning

Abstract

This paper proposes an algorithm for automatic segmentation of video clips into speaker units, with the intention of using the latter as the index units for a video indexing and retrieval system. The algorithm works by pooling together global information of each speaker before detecting the true speaker change locations. The use of stacked scanning can offer better results. The experiment shows that the algorithm is able to boost the discriminating power of BIC resulting in 20% reduction in average detection offset and 5% reduction in average maximum detection offset.