The Impact of the Threshold Values on the CNN Partition Approach in the VVC Standard

In 2020, the joint video expert team (JVET) developed a new video compression standard named Versatile Video Coding (VVC). In addition to High Efficiency Video Coding (HEVC) tools, the VVC scheme has added new coding techniques such as the Quadtree with Nested Multi-Type Tree (QTMT) partition. The QTMT has improved the coding performance while increasing computational complexity by adding Binary Tree (BT) and Ternary Tree (TT) partitioning structures. Therefore, we proposed a fast QTMT decision tree for VVC based on a deep neural network. Our algorithm implemented a Convolution Neural Network (CNN) to predict the BT splitting, where threshold values were adopted in order to achieve a trade-off between the complexity reduction and the coding performance. In this paper, based on our published approach, we realize a study of how the threshold values were chosen at different levels of CU partition and how they affect the RD performance and the coding complexity. This study aims to propose an algorithm that can select the best threshold values.