JVET-AE0104 AHG10: GOP-based RPR encoder control using parallel resolution encoding [D. Arai, S. Nemoto, S. Iwamura, A. Ichigaya (NHK)]

This contribution aims to improve the performance of JVET-AB0080 in controlling the encoder for GOP-based Reference Picture Resampling (RPR). The GOP-based selection method utilizes parallel resolution encoding for Rate-Distortion (RD) optimization. This method performs multi-pass encoding on the downscaled pictures and then calculates the RD cost for each resolution to determine the optimal resolution in each GOP.

The BD-rate performance (Y/Cb/Cr) compared to CTC with VTM-20.2 as anchor is as follows:

• RA: -0.65%/1.52%/0.48%
• LDB: -0.06%/-0.12%/-0.33%
• AI: -0.21%/4.83%/2.23%

This method is especially beneficial for 4K resolution in the RA configuration (-2.24%/-0.18%/-2.90% for ClassA1, -1.14%/7.42%/5.36% for ClassA2).

The method also improves over current GOP-based RPR (except for AI, where gains are experienced in some classes); this also comes with a run time increase (50% in RA, 86% in LB).

It was noted that in AI this is not using RPR in the loop, it is using VVC interpolation filters for upsampling.

As RD optimization is used per RPR including I picture, parallel encoding of GOPs in RA is not possible.

It was commented that weights of 8:1:1 for RD cost calculation may not be appropriate.

Further study was recommended along the questions asked above.