JVET-Q0162 A memory issue on adaptive colour transform [K. Kondo, M. Ikeda, T. Suzuki (Sony)]
In this contribution, it is proposed to restrict the large transform such 64-points when an adaptive colour transform (ACT) is used. Also, it is introduced that ACT needs large buffering because the inverse-ACT needs three colour components after inverse transform. To reduce maximum buffer size, it is proposed that sps_act_enabled_flag is implicitly set equal to zero when sps_max_luma_transform_size_64_flag is true. The test is carried out and report the impact for coding efficiency. Based on the same test as JVET-P0517, simulation results reportedly show that when the separate-tree partition is off, the average {G, B, R} BD-rate losses {0.3%, 0.1%, 0.2%}, {0.4%, 0.1%, 0.3%} and { 0.5%, 0.2%, 0.3%} for AI, RA and LD configurations, respectively.
It is agreed that the issue raised here is valid, and that the loss observed by disabling 64-point transform when ACT is enabled is acceptable (considering the gain that ACT gives for RGB coding). It is pointed out that other solutions could be possible:
- Formulate as bitstream restriction at SPS
- Implicit split to 32x32 when ACT is enabled
- Disable ACT for CU sizes >1024 samples (proponents say they studied this approach, but loss was larger)
Decision: Adopt JVET-Q0162, disable ACT when max transform size is 64 (CTC for 4:4:4 RGB shall use 32 max transform size and ACT).