JVET-AM0227 AhG7: On bin to bit ratio in ECM [T. N. Canh, P. Yin, S. McCarthy (Dolby), J. N. Shingala (Ittiam)] [late]
This contribution studies CABAC bin statistics across ECM 17.0, VTM 23.10, and HM 16.19 under CTC condition. The peak and average bin to bit ratio (bin2bit), weighted BD-binrate, and context coded ratio are examined. For weighted bin2bit calculation, bypass bins are weighted at 0.25. The results for AI/RA in CTC condition are as follows:
- ECM 17.0 shows 17.38%/16.40% (7.81%/6.62%) increase for peak weighted (unweighted) bin2bit ratio over VTM 23.10 and 41.31%/40.35% (22.11%/22.03%) over HM 16.19, while VTM 23.10 shows 20.39%/20.58% (13.52%/12.58%) increase for peak weighted (unweighted) bin2bit ratio over HM 16.19.
- ECM 17.0 shows weighted BD-binrate of -2.22%/-11.12% over VTM 23.10 and weighted BD-binrate of -11.70%/-40.97% over HM-16.19 for luma while VTM 23.10 shows weighted BD-binrate of -10.11%/-33.53% for luma over HM 16.19.
- The ratio of Y bitrate saving to Y binrate saving of ECM-17.0 over VTM-23.10 is 7.59x/2.43x, while the corresponding ratio of VTM23.10 over HM 16.19 is 2.53x/1.27x.
- ECM 17.0 has a context bin over total bin ratio of 86.98%/85.68% which represents an increase of 13.68%/13.72% from 76.51%/75.34% in VTM 23.10 and 24.62%/25.31% over HM 16.19 from 69.79%/68.37%.
- The number of contexts has increased by a factor of 4.5x from 379 contexts in VTM 23.10 to 1706 contexts in ECM 17.0. The storage for CABAC initialization and context models has increased by a factor of 22.5x and 10.8x in ECM 17.0 over VTM 23.10, respectively.
The ECM requires a higher throughput CABAC engine for real time decoding due to double digit increase in bin2bit ratio from VTM. The peak weighted bin2bit ratio has been increased from 1.1 (1.02 ~ 1.16) in VTM to 1.3 (1.20 ~ 1.54) in ECM. On the other hand, the binrate saving is much smaller in ECM over VTM compared to VTM over HM due to the significant increase in signalling bins from newly added. Meanwhile, most bins are context coded in ECM at nearly 86~87% of total bins. Additionally, the memory requirement for CABAC initialization and context models has been increased significantly in ECM 17.0 over VTM 23.10.
Reference is taken to JVET-N0049, where the bin-to-bit ratio is introduced as a criterion for throughput. Number of context models is suggested as another relevant aspect, having impact on storage needs. Also the weighting factor comes from JVET-N0049.
It was commented that it is less relevant for an exploration item of ECM to have an excessive usage of context coded bins, but for a potential proposal towards a future standard it is highly important.
It is further proposed in the contribution to fix a bug in VTM (and ECM), which does not allow to analyse the bin to bit ratio at the current moment. It was suggested to submit a report of the bug and a merge request.