JVET-O0194 On chroma-CU-related flags and coding information [Y. Zhao, H. Yang (Huawei)]

When separate tree is used, the coding information CuPredMode[x][y], cqtDepth[x][y], CbWidth[x][y], CbHeight[x][y], CbPosX[x][y] and CbPosY[x][y] of location (x,y) may be different for luma and chroma CBs. However, VVC draft 5 does not discern these parameters for luma and chroma. To fix this problem, it is proposed to extend these parameters with channel type in VVC spec, e.g., CuPredMode[chType][x][y]. Besides, the first four of the six parameters are used in the neighbour-based context modelling of pred_mode_ibc_flag, split_qt_flag, mtt_split_cu_vertical_flag and split_cu_flag, thus requiring line buffer to store these parameters for luma and chroma separately.

Three tests are made to remove neighbour dependency in context modelling of the four flags for chroma CU, by which the line buffer for storing CuPredMode, cqtDepth and CbWidth for chroma blocks is reduced. Test 1 does not check neighbours for the first two flags for chroma CU. Test 2 does not check neighbours for all the four flags for chroma CU. Test 3 does not check neighbours for the first three flags for chroma CU, and checks only the left neighbour for the split_cu_flag. It is reported that Test 1 results in 0.00%/-0.01%/0.02% and 0.02%/-0.03%/0.01% Y/Cb/Cr BD-rates in AI and RA configuration, respectively. Test 2 results in 0.06%/0.23%/0.25% and 0.02%/-0.02%/0.11% Y/Cb/Cr BD-rates in AI and RA configuration, respectively. Test 3 results in 0.02%/0.13%/0.22% and 0.00%/0.00%/-0.03% Y/Cb/Cr BD-rates in AI and RA configuration, respectively, and saves 6 bits/4x4 in line buffer.

The contribution points out some issues for the case of dual tree luma/chroma dependency.

First aspect: CB width/height is purely editorial (no change in software)

Second aspect: remove context modelling from any neighbours for chroma, loss is 0.06% in luma, 0.2% in chroma

Third aspect: remove context modelling only for neighbours above, saves line buffer for context model (about 5 bit per 4x4 block, approx. 5000 bit for case of 4K), loss 0.02% in luma,0.1-0.2% in chroma.

Question is raised: Is this really a problem?

No support by other experts – no action on aspects 2 or 3.

Decision (Ed.): Adopt JVET-O0194 first aspect – editorial improvement.