JVET-N0194 CE6/7-related: Context selection of last non-zero coefficient position coding based on reduced TU size (related to JVET-M0297 and JVET-M0251/M0257) [M. Koo, J. Choi, M. Salehifar, J. Lim, S. Kim (LGE)]
In 13th JVET meeting, reduced 32-point MTS (RMTS32) based on skipping high frequency coefficients was adopted in order to reduce computational complexity of 32-point DST-7/DCT-8. Moreover, it was also agreed that the proposed residual coding change should be accompanied considering all types of zero-out (those of JVET-M0297 and JVET-M0251/M0257).
During in-between VVC working draft editing period, all the related proponents provided their spec text where contexts of last non-zero coefficient position coding is selected based on original TU size, which is aligned with the current VTM4.0 SW. But, it was questioned if this context selection should be performed based on reduced TU size, which was commented to be more consistent. Therefore, it was proposed that it could be decided at the next meeting based on experimental results which TU size should be involved during the context selection.
In this contribution, the experimental results based on reduced TU size are presented. It is reportedly shown that luma BD-rate/encoding time/decoding time changes of all configurations are 0.10%/101%/103% (AI), 0.05%/XXX%/XXX% (RA), and -0.02%/XXX%/XXX% (LD), respectively, but BD-rate loss for high resolution sequences on AI configuration is substantial (up-to 0.24%).
The presentation deck was not available in latest upload and was requested to be made available.
Decision (BF): The problem that this contribution raises is recognized. The text should be aligned with the method imlemented in VTM, as the latter seems to have better performance (to be done by text editor).