Re-Engineering Rate Distortion Optimisation in Modern Video Codecs Using a Per Clip Approach

Abstract

The majority of internet traffic is video content. This drives the demand for video compression to deliver high quality video at low target bitrates. Optimising the parameters of a video codec for a specific video clip (per-clip optimisation) has been shown to yield significant bitrate savings. One of the key challenges to be solved in the design of a practical codec is the trade-off between rate and distortion. A Lagrangian multiplier has been the adopted view in video codecs. In this approach the encoder makes a number of decisions in order to minimise a cost J, where J = D + ? R, commonly referred to as the rate distortion equation. As can be seen J combines both a distortion D (for a frame or macroblock) and a rate R (the number of coded bits for that unit) through the action of the Lagrangian multiplier ? . Different choices for ? result in different R/D trade-offs. This research investigates the impact of adjusting the rate distortion equation on com- pression performance. A constant of proportionality, k, is used to modify the Lagrangian multiplier used in H.265 (HEVC) and VP9. Direct optimisation methods are deployed to maximise BD-Rate improvement for a particular clip. By doing this, we have shown that per-clip optimisation of the Lagrangian multiplier leads to BD-Rate improvement of up to 25% in a single clip using Direct Optimisation. The average BD-Rate improvement was approximately 2% across the 10k clip corpus. This direct optimisation comes at a high computational cost, but can be used as an estimate of the best possible improvement possible for a given video clip. Because of this computationally expensive method which requires multiple measurement of rate distortion curves which meant in excess of fifty video encodes were used to generate that level of savings. This research then focuses on reducing the computational cost of repeated video encodes by using proxy systems or by reducing this computational load through prediction. Approximately 60% of the BD-Rate improvements found by direct optimisation using proxies or prediction were achieved using the reduced complexity systems. This effectively reduced the computational complexity ten fold. Overall, our system achieves BD-Rate improvement in approximately 90% of a large corpus with comparable results to the direct optimisation methods.