JVET-N0049 AHG16: A study of bin to bit ratio in VTM4.0 and HM16.19 [M. Zhou (Broadcom)]
It is observed that the bin to bit ratio of the VTM4.0 is double digits percentage higher than that of the HM16.19, for both the CTC and low QP coding (QP =2, 7, 12, 17). The summary results are as follows:
- For CTC AI/RA configurations, on average the weighted bin to bit ratio is 15.72%/17.39% higher than the HM16.19, and the unweighted bin to bit ratio is 12.38%/13.75% higher than the HM16.19, respectively.
- For low-QP AI/RA configurations, the weighted bin to bit ratio is 18.11%/17.17% higher than the HM16.19, and the unweighted bin to bit ratio is 10.45%/7.63% higher than the HM16.19, respectively.
The BD-rate reduction of the VTM4.0 relative to the HM16.19 at low QPs (QP =2, 7, 12, 17) is -12.39%, -13.99%, -13.69% and -14.69% in AI, RA, LD_B and LD_P configurations, respectively.
In the weighted bin to bit ratio, a bypass bin is counted at 0.25 context coded bins; in the unweight bin to bit ratio, a bypass bin and a context coded bin carry an equal weight (1:1).
This was presented Tuesday 03-26 1720 (chaired by JRO)
The contribution points out that a relatively higher percentage of bins seems to be context coded rather than bypass coded. A calculation is made that weights context coded bins by a factor of 4, and bypass coded bins by a factor of 1, to reflect the higher complexity of contex coded bins. The resulting “BD rate” based on the distortion-complexity criterion indicates that VVC is worse than HEVC.
Further study is necessary. It is however well known that the ratio between complexity and additional compression is not linear. It would in particular be interesting how many context coded bins are consumed by which part of the bitstream, such as transform coefficients, mode/motion etc.