Organisation internationale de normalisation
SCC complexity reduction (AHG9&10) (3)
Yüklə 9,04 Mb.
|
SCC complexity reduction (AHG9&10) (3)13.0.0.1.1.1.1.1.308JCTVC-S0068 AHG14: On IBC memory reduction [G. Laroche, T. Poirier, C. Gisquet, P. Onno (Canon)] (Consideration of this topic was chaired by G. Sullivan on Saturday 10-18 p.m.) An IBC PU predictor comes from reconstructed non-filtered blocks. This increases the memory needed for some implementations, an example being the case where loop filtered blocks need to be stored in addition to the reconstructed blocks because of IBC. To avoid this additional storage, this contribution proposes to signal which CTUs are available for IBC prediction. For these CTUs, the DBF and SAO are disabled. The proposed method reportedly gives an average of 0.7% loss for the AI configuration compared to the current SCM2.0. This proposes to code a CTU-level flag, available_for_ibc_flag, to identify whether a block may be referenced by IBC. When equal to 1, DBF and SAO are disabled within that CTU. (On other side of an edge, there may be filtering.) AI average loss is average 0.7%. About 2% average loss for RA & LD. This seems excessive. The peak category of loss was YUV 1080p mixed content, which had 5.1% average loss. It was remarked that some kind of modified encoding search could probably reduce that loss. Further study of this was encouraged. 13.0.0.1.1.1.1.1.309JCTVC-S0241 Cross-check report of JCTVC-S0068 On IBC memory reduction [K. Rapaka (Qualcomm)] [late]
(Consideration of this topic was chaired by G. Sullivan on Saturday 10-18 p.m.) This contribution proposes a method to reduce average bandwidth when intra block copy (IBC) mode is used for prediction. IBC mode uses previously decoded unfiltered samples within the same picture for its prediction. In the current test model, for the IBC mode the search range is unrestricted and can use any unfiltered decoded samples of the current picture that lie in previously decoded regions. This introduces additional bandwidth required to store unfiltered decoded samples of the current picture (in addition to filtered decoded samples required for temporal prediction). It is observed that on average not all previously decoded unfiltered samples of the current picture are used for prediction in IBC mode and the usage depends on the characteristics of content, for example, the usage peaks for text and graphic sequences and is minimal for camera-captured content. Hence, it is argued that always storing the previously decoded unfiltered samples for the current picture is inefficient from the bandwidth perspective. In this contribution it is proposed to indicate which of the previously decoded CTUs are used for IBC prediction so that only those CTBs could be stored thereby reducing the average bandwidth. Similar to S0068, the contribution proposes to send one flag per CTU to identify which parts of the picture need to be additionally stored for non-filtered referencing. However, it does not disable the ILFs and does not reduce the worst case. It was remarked that the write function is not as much of a problem as the readback. The implications of needing to know the how many CTBs are in the slice and needing to send the flags in advance were questioned. The proponent indicated that it would be possible to send this information as an SEI message. It was asked whether IBC is constrained to reference only within a slice and/or within a tile. An editor said he thought such constraints exist, but would doublecheck. Further study of this was encouraged.
|