Complexity reduction methods for Versatile Video Coding: A comparative review

To deal with the growing demand for ultra-high definition video and increasingly challenging requirements of services and applications, the Versatile Video Coding (VVC) is the most recent standard, significantly improving the coding efficiency in comparison with its predecessors. However, since such improvement is obtained at the cost of a great increase in computational complexity, worldwide research has been underway to find methods capable of reducing the VVC complexity without compromising its coding efficiency. Given the number and diversity of the methods presented in the literature in recent years, every new research only analyses and compares with a limited number of previous works, very often selected without clear criteria. Furthermore, we have found that the usual way to assess and compare complexity reduction methods, based on the tradeoff between complexity gain (in percentage) and coding efficiency loss (in Bj & oslash;ntgaard delta rate (BD-Rate)), fails to be a valid performance indicator across the whole operational range. This paper is a contribution to establish a research reference by presenting a comprehensive comparative review of methods specifically proposed for complexity reduction of VVC. Two new performance comparison metrics are also proposed using the complexity gain and BD-Rate loss as parameters. One of them is characterised by a linear behaviour and the other based on the distance to an efficiency frontier defined by the maximum complexity gain for a given BD-Rate loss. This comparative study takes into account different versions of the software implementation through a normalisation approach, which allows fair comparison of different methods implemented on different encoder software versions. In general, it is shown that Machine Learning (ML) based methods usually outperform heuristic ones while fast methods for intra coding mode estimation present the highest complexity reduction opportunities. Overall this paper provides a novel comparative study of 83 methods and proposes fair performance comparison metrics that are useful for further research in the field and also for future developments on hybrid video coding approaches with reduced computational complexity.