JVET-P0538 Non-CE6: Transform skip with fixed transform shift [J. Xu, L. Zhang, W. Zhu, K. Zhang, H. Liu, Z. Deng (Bytedance)]
In the current VVC design, a transform shift is applied to simulate transform process for all blocks even for a transform skip coded block. In this contribution, two methods are proposed. Method 1 is to have a fixed shifting to fit the 16-bit range. Method 2 is to remove such a transform simulation for transform skip blocks. It is asserted that computational complexity reduction is achieved. No BD-rate changes under the CTC are observed. For low QPs (2, 7, 12, 17), some minor BD-rate changes are reported.
The deviation is due to the fact that the current method of VVC invokes a right shift (block size dependent) in dequantization which introduces a kind of rounding error that cannot be compensated by the subsequent left shift. This effect becomes noticeable only at very low QPs.
The proposed methods solve this issue and by tendency have (marginally) better performance, in most cases bit-wise equal. Method 2 which is basically performing “transquant-bypass with quantization/scaling”
JVET-P0272 and JVET-P0515 are targeting the same issue
JVET-P0272 and JVET-P0515 are identical to method 2 at decoder, but have a modified encoder in terms of RD optimization. The additional gain is not high (0.1% in TGM and class F in AI for JVET-P0515)
Proponents of the three contributions to come together, and provide text, software and a recommendation of possible changes to the encoder. See JVET-P1000.