TU Delft at TRECVID 2005: Shot Boundary Detection

In TRECVID 2005 we tested a simple but effective method for simultaneously detecting video shot transitions of various types by means of an analysis of spatiotemporal video data blocks. This method differs from the existing approaches in the way that it takes volumetric data cubes in the video as the fundamental processing unit (sysID: TUDelft1 and TUDelft2). TUDelft1 also involves an illumination and flash light normalization component, which resulted in considerable amount of increase in overall performance figures. Due to its simplicity, the proposed method is highly computationally efficient. We observed that the most important problem of the system is its fragility against motion in gradual transition detection. This stems from the fact that the analysis currently takes the gradient information only in time direction. We plan to correct for this problem in 2006 when we updated the system also for camera motion detection.