Back to Search Document details
17th Meeting: Brussels, January 2020 2020-01-11 08:10
Non-CE2: On signaling of maximum palette size and maximum palette predictor size
Authors: Xiaoyu Xiu Kwai Hong-Jheng Jhu Kwai Yi-Wen Chen Kwai T.-C. Ma Kwai X. Wang Kwai
Abstract
In HEVC standard, maximum palette table size and maximum palette predictor size are signaled in sequence parameter set (SPS) while the palette in the VVC draft 7 uses fixed maximum palette table size of 31 and fixed maximum palette predictor size of 63. In this contribution, it is proposed to bring the syntax of signaling maximum palette table size and maximum palette predictor size in the HEVC back to the palette design in the VVC. The proposed signaling was implemented on top of VTM-7.0 and tested using different maximum palette table size and maximum palette predictor size. Compared to VTM-7.0 anchors, the performance for the sequences in the category YCbCr444 TGM 1080p is summarized as
Meeting Notes
JVET-Q0519 Non-CE2: On signalling of maximum palette size and maximum palette predictor size [X. Xiu, H.-J. Jhu, Y.-W. Chen, T.-C. Ma, X. Wang (Kwai)]

In HEVC standard, maximum palette table size and maximum palette predictor size are signalled in sequence parameter set (SPS) while the palette in the VVC draft 7 uses fixed maximum palette table size of 31 and fixed maximum palette predictor size of 63. In this contribution, it is proposed to bring the syntax of signalling maximum palette table size and maximum palette predictor size in the HEVC back to the palette design in the VVC. The proposed signalling was implemented on top of VTM-7.0 and tested using different maximum palette table size and maximum palette predictor size. Compared to VTM-7.0 anchors, the performance for the sequences in the category YCbCr 4:4:4 TGM 1080p is summarized as

For maximum palette table size of 64 and maximum palette predictor size of 128, for lossy coding, it provides average {Y, Cb, Cr} BD-rate differences of {-0.66%, -1.23%, -0.96%} and {-0.43%, -0.79%, -0.69%} for AI and RA configurations, respectively. For lossless coding, the average AI and RA bit-rate differences are -0.47% and -0.36%.

For maximum palette table size of 32 and maximum palette predictor size of 64, for lossy coding, it provides average {Y, Cb, Cr} BD-rate differences of {2.00%, 2.95%, 2.59%} and {1.46%, 1.84%, 1.73%} for AI and RA configurations, respectively. For lossless coding, the average AI and RA bit-rate differences are 2.30% and 2.17%.

Question: Could that also be achieved by an encoder-only method (e.g. an encoder only using first 32 entries)?

It is not proposed to increase the maximum of the current spec (31 for palette, 63 for predictor). The results with larger tables are only given for information.

The proposal would mainly for the benefit of encoders. Software decoders might benefit having less power consumption when knowing that the table is smaller.

It is shown that the number of checks per CU for constructing the palette table and the predictor table is reduced by half when the configuration PLT16/Pred32 is used. It is not clear from the information given how much reduction of encoder/decoder run time would be achieved by this configurability.

References:
JVET-P2001 JVET-P2022 JVET-P2023
Cited:
JVET-Q0011 JVET-Q0022 X JVET-Q0716 X
PATENTS:
4088455 0.34 3146722 0.33 20180091829 0.32 3243330 0.31 WO/2016/127889 0.30 WO/2021/138432 0.28 3117608 0.27 20220353540 0.26
ARCHIVE v2 :
JVET-P0519.pptx JVET-Q0519.docx JVET_Q0519_pltSize16_predSize32_LL.xlsm JVET_Q0519_pltSize16_predSize32_LY.xlsm JVET_Q0519_pltSize64_predSize128_LL.xlsm JVET_Q0519_pltSize64_predSize128_LY.xlsm
Decisions
It is shown that the number of checks per CU for constructing the palette table and the predictor table is reduced by half when the configuration PLT16/Pred32 is used. It is not clear from the information given how much reduction of encoder/decoder run time would be achieved by this configurability.
Citation