The Dynamic Optimizer Framework. Video encoding, assessment and comparison

Delivery costs, low latency, and streaming at scale are key challenges in driving the technical development of the video industry. Improvements in video coding are always pushing the effort for new video coding solutions and quality monitoring. The recently released coding standards, like AV1 and VVC, are proof of this. In line with the constant aim of ensuring better compression ratios for the same visual quality, we analysed a new approach to digital video compression. The so-called Dynamic Optimizer has been proposed and conceived originally by Netflix, it is compatible with any existing and future video codecs, and it is suitable for non-real time encoding of on-demand video contents in adaptive streaming applications. Therefore, its ideal resources are long sequences of shots, typical of the Netflix catalogue. It works by fine-tuning the available encoding parameters for the single shots in order to find the optimal combination that respects a given bitrate or quality target. We developed a software capable of reproducing the proposed optimizer, which will be released with an open-source licence. It consists of three alternative implementations: a brute force approach that encodes and compares all possible combinations and it is very expensive in terms of number of computations and encodings, and the Lagrangian optimization, based on searching the optimal minimum of the convex hull of all combinations. The third and the most efficient proposed solution instead, which constitutes the innovative element of the work, is based on the RD curves approximation. In this thesis, we analyse the performances of the dynamic optimizer thanks to a comprehensive set of objective assessment tests, using the VMAF quality metric and the AVC, HEVC, VP9, and AV1 encoders. The results highlight the advantages and disadvantages of the different encodings in various bitrate and quality ranges and for diverse types of content.