JVET-R0378 TU split for ACT [K. Kondo, M. Ikeda, T. Suzuki (Sony)]
In this contribution, it is proposed to allow a large transform such 64-points when an adaptive color transform (ACT) is disabled at the coding block. In the last meeting, to restrict the large transform was adopted. It can help to reduce buffer size and latency between inverse-transform and inverse-colour transform. However, it was restrictive because the 64-points transform can not be used even if ACT was disabled. This contribution reports the test results when the 64-points transform is allowed. Based on the common test condition for RGB 4:4:4 as described in JVET-Q2013, simulation results reportedly show that the average {G, B, R} BD-rate {0.01%,-0.26%,-0.25%} for screen contents and {-0.57%, -0.25%, -0.43%} for natural contents in AI configurations.
Conceptually similar with JVET-R0305, but the latter has somewhat more gain (where it was however questioned how these relate to the bug in previous software). JVET-R0305 additionally applies the adaptive transform size restriction to subblocks.
V3 of JVET-R0378 shows additional results which correct the settings in terms of QP table as suggested in JVET-R0321. Overall gain is 0.25%/0.25%/0.29% in 4:4:4 RGB CTC for AI, and for natural content it is 0.6%/0.26%/0.39%, which seems to be similar to the results reported in the abstract.
There are additional results for RA in the new version, which are still incomplete.
It is questioned how relevant a use case is of coding RGB 4:4:4 content in a lossy mode (with CTC up to QP37), when it could be converted to YUV, which is usually saving several percents. Compared to that, the gain in AI is relatively small, and for screen content it is almost zero. This does not justify a low-level change.
No action was taken on this.