Detection and Video Search

In this paper, we describe our methodologies and empirical evaluations for the shot boundary detection and automatic video search tasks at TRECVID 2006. For the shot boundary detection task, we consider a simple and efficient solution. Our approach first applies adaptive thresholding on color histogram differences between frames to select candidates for shot boundaries, then runs several tests to dismiss less likely ones. These tests proved to be too finely tuned to the TRECVID 2005 shot boundary detection task data, producing mediocre results on the 2006 data set. For the video search task, we propose a novel multimodal and multilevel ranking scheme for video ranking. Different from traditional ranking schemes, most of them are based on supervised approaches, our approach suggests a semi-supervised ranking method which can exploit both labeled and unlabeled data effectively for the ranking task. At the meantime, our multilevel approach makes the semi-supervised ranking method efficient for large-scale search task in practice. We will evaluate the empirical performance of our approaches and give comments of our solution. 1