Our approach is based on feature tracking where cuts are detected when only few features can be tracked across frame pairs. This unique criterion was found sufficient to detect most cuts in most videos even in the case of short term transitions. Furthermore, we propose a method to stabilize the differences and automatically identify a global threshold to achieve a high detection rate. For the TRECVID06 submitted runs, we tested the algorithm under the following conditions: • CU-UO-All-Runtype:1 corresponds to the Precision set as determined by automatic threshold selection; • CU-UO-All-Runtype:2 corresponds to the F1 set as determined by automatic threshold selection; • CU-UO-All-Runtype:3 corresponds to the Recall set as determined by automatic threshold selection; • The other runs correspond to various manually selected threshold values. One important aspect of our system resides in the user control ability over the type of acceptable errors; according to this, the system will select automatically the most appropriate threshold to minimize the specified type of error (false negatives versus false positives). From the experiments, this component works as expected even though our results were less effective than expected. The exact reason for the exhibited performance is still to be determined currently being investigated. The method was initially designed for the segmentation media types other than news (e.g. movies, television shows, cartoons, etc.). Within the news media, the existence of graphical overlays introduces a number of unique issues that includes having to adjust the feature detection to areas where the video are changing. Indeed, computer graphics introduce sharp textures where more features are detected. This also brings us to the definition of a cut in the case where there is a mix of natural background videos and graphical overlays that are changing at different rhythms. In cinema and television it is common to have few frame shots (even just 1 or 2 frames), that situation does not occur in news so having a wider temporal search range. However, while doing so would improve the results for news, this might result in missed shot in other media types.

Additional Metadata
Conference TREC Video Retrieval Evaluation, TRECVID 2006
Whitehead, A, Bose, P, & Laganière, R. (Robert). (2006). TRECVID 2006 notebook paper: Feature based cut detection with automatic threshold selection. Presented at the TREC Video Retrieval Evaluation, TRECVID 2006.