On the performance of fast motion estimation for multiple-reference motion compensated prediction

Multiple reference motion compensated prediction achieves significant prediction gains, but at the expense of a significant increase in computational complexity. This paper investigates and compares the performance of a number of fast motion estimation techniques when extended to work for multiple reference prediction. In general, algorithms that are designed to exploit the properties of the long-term block motion field, were found to provide superior performance and better compromise between prediction quality and computational complexity.