Research

Fast encoding parameter selection technique have been proposed in the past. Leveraging the power of convex hull video encoding framework, an encoder with a faster speed setting, or faster encoder such as hardware encoder, can be used to determine the optimal encoding parameters. It has been shown that one can speed up 3 to 5 times, while still achieve 20-40% BD-rate savings, compared to the traditional fixed-QP encodings. Such approach presents two problems. First, although the speedup is impressive, there is still ∼3% loss in BD-rate. Secondly, the previous approach only works with encoders implementing standards that have the same quantization scheme and QP range, such as between VP9 and AV1, and not in the scenario where one might want to use a much faster H.264 encoder, to determine and predict the encoding parameters for VP9 or AV1 encoder. In this work we propose and present new additions to address these issues. First, we show that the predictive model can be used to gain back the quality loss incurred from using a faster encoder. Secondly, we also demonstrate that, by introducing a mapping layer, it is possible to predict between two arbitrary encoders and reduce the computational complexity even further. Such degree of complexity reduction is made possible by the fact that different generations of video codecs are usually orders of magnitude apart in terms of complexity and thus one can use an encoder from an earlier generation to predict the encoding parameters for an encoder from later generations.