Real-time Estimation of Subjective Utility Functions for MPEG-4 Video Objects

In this paper, we propose an original system for a subjective utility function estimation suitable for real-time video applications. In video networking applications, utility function is deened as a video quality metrics representing a user satisfaction index as a function of bandwidth. The real-time performance is achieved in our system by (1) using novel content classiication techniques based on visual features directly extracted from compressed video streams and (2) employing machine learning techniques for classiication of utility functions. Our extensive experiments, based on MPEG-4 encoded video streams, indicate that high accuracy of 80%-85% in estimating utility functions can be achieved. These results demonstrate that the proposed system represents a promising approach to real-time utility function estimation. The real-time performance is instrumental in achieving higher network resource utilization leading to exible and less expensive interactive multimedia services. The proposed technique is general and can be applied to other video compression standards such as MPEG-1, MPEG-2, and H.263. An important application of the proposed system is in future low bandwidth and wireless networks that can support and beneet from content-aware media scalability.