FXPAL at TRECVID 2006

The shot boundary detection system we are using for 2006 builds on the framework and system developed in 2004 and 2005 which combines pairwise similarity analysis and supervised classification. Using primitive low-level image features, we build secondary features based on inter-frame dissimilarity. These secondary features are used as input to an efficient k-Nearest-Neighbor (kNN) classifier. The classifier labels each frame as a shot boundary or non-boundary, and the classifier outputs are minimally processed to determine the final segmentation. Last year our performance was worse than anticipated, as our training data was not an accurate reflection of the test data for the videos from LBC and CCTV. On the remaining videos, our performance was very good, and consistent with our training experiments. As a result, this year we submitted experimental runs using machine generated output from the master shot reference of the development set to label the training data used for classification. Our results were reasonably good, but remained below expectations and previous performance using manually generated training sets. Further analysis revealed that errors in our frame decoding software were responsible for most of this performance loss.