JVET-L0381 CE3-related: 4-tap interpolation filter selection with quantization parameter [Y. Kidani, K. Kawamura, K. Unno, S. Naito (KDDI)]

This contribution proposes 4-tap cubic and Gaussian interpolation filter selection with quantization parameter (QP) which is related to Core Experiment 3 (CE3). These filters are applied to JEM-7.0 and are studied in CE3.3 and which filter should be used is selected with intra block sizes and intra directional modes. The thresholds of block size to separate the usage of both filters is kept constant at 8 width or height even though QP values change. In this contribution, the thresholds are proposed to be optimized by versatile QP ranges and the simulation results for this technique reportedly show -0.40%/-0.14%/-0.15% BD-rate for luma gains in AI/RA/LDB, respectively, in the CTC QP range compared to VTM2.0.1 where 2-tap interpolation is applied. Moreover, from the simulation results by versatile QP range, it is confirmed that at high QP range BD-rate for luma in AI increase -0.17% compared to the results with conventional block-sized-based selected filters.

The base proposal CE3.1.1 determines the switching by block and intra mode. Here, the QP is used as additional criterion. Under CTC, the additional gain is around 0.04%. In QP range 32-47, the gain the proposal shows on top of CE3.1.1 becomes larger on average (0.23%), however for class A, where such QP ranges might be applied, there is no gain (even loss for A1). No action was taken on this.