Generalized bi-prediction method for future video coding

This paper presents a generalized bi-prediction (GBi) technique to extend the notion of the existing weighted bi-prediction to prediction unit (PU) level. GBi allows bi-prediction weights to be transmitted at the coding unit (CU) level and used for the bi-predicted PUs within that CU. The candidate set of weights in the GBi mode includes a total of 7 weights, including 1/2 used for conventional bi-prediction. An index pointing to the entry of a weight value in the candidate weight set is signaled. At most one index per CU is signaled and the corresponding weight values are shared across all the bi-prediction PUs and all color components in that CU. Besides, GBi can also be extended to inter coding tools, such as, merge mode and affine prediction that support bi-prediction. Experimental results show that GBi performs consistently better than the JEM-2.0 anchor in all rate points, with an average Y Bjentegaard delta rate reduction of 1% and no extra decoding burden introduced.