Back to Search Document details
23rd Meeting: by teleconference, July 2021 2021-07-08 21:15
AHG10: GOP-based RPR encoder control
Authors: K. Andersson Ericsson J. Ström Ericsson Ruoyang Yu Ericsson
Abstract
VTM currently lacks encoder control for selecting when to use Reference Picture Resampling (RPR). This contribution proposes a GOP-based selection method that decides when to encode pictures in reduced resolution with RPR or if its better to encode them in the source resolution. The selection is based on QP and picture self-similarity after re-scaling. The self-similarity test is only conducted on the first source picture in display order within each GOP. If the re-scaled picture is determined to have sufficient similarity with the source picture, all pictures in the GOP are encoded at reduced resolution, otherwise they are encoded in the source resolution.
Meeting Notes
JVET-W0082 AHG10: GOP-based RPR encoder control [K. Andersson, J. Ström, R. Yu (Ericsson)]

The VTM currently lacks encoder control for selecting when to use Reference Picture Resampling (RPR). This contribution proposes a GOP-based selection method that decides when to encode pictures in reduced resolution with RPR or if its better to encode them in the source resolution. The selection is based on QP and picture self-similarity after re-scaling. The self-similarity test is only conducted on the first source picture in display order within each GOP. If the re-scaled picture is determined to have sufficient similarity with the source picture, all pictures in the GOP are encoded at reduced resolution, otherwise they are encoded in the source resolution.

The BD-rate performance (Y/Cb/Cr) compared to CTC with VTM-13.0 as anchor is as follows:

RA (luma/Cb/Cr): -0.03%/0.61%/0.58%

LDB (luma/Cb/Cr): 0.27%/0.35%/0.35%

AI (luma/Cb/Cr): 0.11%/0.63%/0.67%

The BD-rate performance for QP 27 to 47 with VTM-13.0 as anchor:

RA (luma/Cb/Cr): -1.57%/6.94%/5.94%

LDB (luma/Cb/Cr): 1.42%/4.06%/3.41%

AI (luma/Cb/Cr): -0.41%/6.39%/4.76%

It is asserted that the method, when it is selected, in some cases can improve subjective quality compared to encoding in source resolution for low bit rates.

The method is only applied for QP = 37 and beyond.

It is pointed out that this might well be combined with segment-wise encoding and open GOP.

It is also pointed out that it mainly seems to be efficient for class A.

It is also mentioned that the effectiveness highly depends on sequence characteristics. For some sequences, gain is achieved only at QP = 50. In this conext, it is recommended to seek for individual “transition points”.

Has it been investigated why there is loss in chroma? See discussion on this elsewhere.

This is a very interesting approach to save bit rate by using coding at lower resolution whenever appropriate. Further study was recommended for further improvements as per the hints above.

It is suggested to also look into subjective quality.

References:
JVET-T2010 JVET-V2016
Cited:
JVET-W0010 JVET-W0173 X JVET-Y0068
PATENTS:
20250343918 0.90 WO/2024/100242 0.90 WO/2025/151060 0.34 WO/2024/210793 0.33 20240340446 0.30 WO/2023/275247 0.28 WO/2023/200381 0.27 WO/2024/080905 0.26
ARCHIVE v3 :
JVET-W0082.docx JVET-W0082_v2.docx JVET-W0082_v3.docx NNVC-1.0_RPR2_config1_QP27_47_v2.xlsm NNVC-1.0_RPR2_config1_v2.xlsm NNVC-1.0_RPR2_config2_QP27_47_v2.xlsm NNVC-1.0_RPR2_config2_v2.xlsm
Decisions
It is suggested to also look into subjective quality.
Citation