Visual Attention Based Video Compression

Accurate prediction of the viewer’s gaze location in a video frame has the potential to improve bit allocation, rate control, error resilience and quality evaluation in video compression. With complex contexts, such as that of broadcast football video, the potential reward is even higher given that compression and transmission of this type of content is challenging. We have developed a gaze location (visual attention) prediction system for high definition broadcast football video. The system employs Bayesian integration of bottom-up features and context specific top-down cues. The context is classified into different categories through shot classification thus allowing our model to pre-learn the task pertinence of each object category and build the top-down prior map automatically.

Q. Cheng, D. Agrafiotis, A. M. Achim, D. R. Bull, “Gaze Location Prediction for Broadcast Football Video”, IEEE Transactions on Image Processing, vol 22, no. 12, pp. 4918-4929, 2013