HuaZhong University of Science and Technology at Trecvid 2006

1、Overview This is the first TRECVID participation for HuaZhong University of Science and Technology . we participated in the Shot Boundary Detection(SBD). In the paper, we will present our approaches for SBD . Our main focus was extracting new features based on the luminance to represent the visual content , and combine the traditional method and sliding windows[7] to build our algorithm. In our algorithm we put forward some post processing modules to make system more robust. From the evaluation results our system shows competitive results. 2、Shot boundary detection Various automatic shot boundary detection algorithms have been proposed ( [5],[6],[9] ). Upon on the recent TRECVID video test set , we find that short gradual transitions and OTHSs become more . Short gradual transitions , OTHS, abrupt illumination change or object/camera’s large movement often lead to significant change of visual content . So in our cut detection algorithm , in order to reduce the disturbance of these things we extract some expressive features , and add post processing modules to process the short gradual transitions and flashlight etc. Detecting gradual transitions is a Challenge work . Our algorithm is based on the Twin comparison method proposed by [4],but it has some shortcomings ,we give a improved twin comparison algorithm. 2.1 Features to measure visual discontinuity Various features have been extracted to represent the visual content of each frame[10]. We extract some simple features based on the luminance , trying to reflect visual content from different aspects. 2.1.1 Improved pixel difference Traditional pixel difference proposed by ([1],[2]) was often effected by noise . To overcoming such shortcomings we give improvements on it . New feature is not sensitive when the whole picture’s luminance become abrupt. In our implementation we compute the pixel difference as: