Joint design of bi-predictive interpolation filters

Motion compensation with fractional pixel accuracy is one of the most effective, yet resource-demanding compression tools of the conventional video coding standards such as HEVC. In this work the design of interpolation filters for bi-predictive motion compensation is revisited to improve the predictive coding gain while reducing the number of filter taps. The proposed filter design uses the sub-pixel motion information from the two reference blocks to derive the interpolation filter coefficients for each of the reference blocks. This is in contrast with the conventional schemes where filter coefficients for a reference block are determined merely by the sub-pixel position information of that block and not the complementary reference block. Experimental results show that compared to the conventional scheme, this technique can be used to increase the coding efficiency for the same number of filter taps or to reduce the number of filter taps while keeping the same coding efficiency.