Improving the temporal segmentation in digital videos using the Deceived Bilateral Filter

This paper presents a novel approach to improve the accuracy of long dissolve detection in the first of the three perception modules of our Automated Sport Analysis System named ACE. Such first module consists in the temporal segmentation of television video broadcasts, decomposing the input digital video into several scenes, based on the scene boundaries found. We propose to include a pre-processing stage for the video signal based in the application of the Deceived Bilateral Filter in order to enhance the dissimilarity obtained between frames during a long dissolve by blurring the frames closest to the long dissolve effect, and increasing image contrast and edges in the rest of the video, in order to achieve better classification results. Results conducted on matches of the last FIFA World Cup 2014 are presented. The classification rates showed precision increase of 14%, a recall improvement of 20% and an overall increase of 28% in the F1-score over the temporal segmentation of the original video.