“…Furthermore, the R-D optimization adopted in DNN-based video compression enables higher compression efficiency by directly using the number of bits in the optimization procedure. To estimate the bit-rates, context models [7], [8], [41], [48], [65] are learned for the adaptive arithmetic coding method which compress discrete-valued data to bit-rates closely approaching the entropy of the representation. More recently, Yang et al [74] shows that hierarchical GOP structure can make use of the advantageous information from high-quality frames by improving the temporal predictive coding.…”