Multiple object labeling in video sequences without skill
Gregory P. House, S. Sakamoto, Junichi Tajima · 2002
The current trend for integrating media of various forms has treated video sequences as object components. Video processing algorithms which automatically extract objects from video sequences need to be fine-tuned to suit the different video contents. The algorithms themselves, however, tend to be too complicated for multimedia-content editors without skills in video processing. We present a framework for use in a multimedia authoring system, that assists unskilled users in labeling objects. Image selection trains pixel- and region-unit feature extraction approaches which were devised to assist editor understanding. Furthermore, a graphical user interface is presented for the tracking stage which enables convenient editor training and label correction. The main part of the framework was implemented on a real-time parallel image processing system. The system was used to label jockeys in a horse racing sequence, achieving a factor of ten time reduction over a comparable manual labeling process.