JVET-R0065 AHG8/AHG9: On IRAP and GDR AUs [Y.-K. Wang (Bytedance)]

This was initially discussedd in AHG Session 1.9 (chaired by GJS & YKW).

Item 22 of this contribution belongs to this category.

This contribution proposes the following changes related to IRAP and GDR AUs:

• Each GDR AU is required to be “complete” (i.e., to have a picture for each of the layers present in the CVS). That means, an incomplete AU consisting of GDR pictures is not a GDR AU, similarly as in the current VVC text that an incomplete AU consisting of IRAP pictures is not an IRAP AU. This aspect seems to be straightforward completion of the specification of GDR to account for multi-layer usage.
• Add a flag named irap_or_gdr_au_flag to the AUD to specify whether the AU is an IRAP or GDR AU, and mandate the presence of an AUD NAL unit in each IRAP or GDR AU when vps_max_layers_minus1 is greater than 0.

The following issues were reportedly observed in the existing scalability design in the latest VVC text (in JVET-Q2001-vE/v15):

• Currently, an IRAP AU may start a new CVS and is required to be complete (i.e., to have a picture for each of the layers present in the CVS), while a GDR AU may also start a new CVS but is NOT required to be complete. It was believed that this is an oversight, as otherwise some layer-wise startup decoding process would have to be specified for such GDR case.
• A CVSS AU, which starts a new CVS, is required to be complete (i.e., to have a picture for each of the layers present in the CVS). However, according to the current design, the decoder is not able to check whether an AU includes a picture "for each of the layers present in the CVS" before it receives the last picture of the CVS, while on the other hand, even the last picture of the CVS is not easy to be determined because it is not easy to determine the start of any of CVS except for the very first CVS of the bitstream. Basically, that means, the decoder can only figure out the boundaries of CVSs after receiving the entire bitstream.

This contribution tries to address the above issues.

An OLS may contain more layers than are present in the bitstream.

It was discussed to what extent there may (or should) be out-of-band signalling to inform the decoder of the start of a new CVS or the establishment of a new target OLS. It was commented that the bitstream should at least be decodable without needing to rely on some such external means.

Another contribution JVET-R0274 proposes to not require an IRAP AU to be complete.

This was further discussed on 23 April at 0845 (chaired by GJS). The purpose of this is to enable early detection of an IRAP or GDR AU, before receiving all pictures in the AU.

The suggestion here is add a flag in the AUD to identify an IRAP or GDR AU when vps_max_layers_minus1 is greater than 0 and mandate the presence of the AUD in these instances.

The flag would not have a normative effect, but would be required to be present under the specified conditions, as a way to assist decoders that would use the flag.

Decision (cleanup): Adopted (item 2 of the proposal).

JVET-R0065 AHG8/AHG9: On IRAP and GDR AUs [Y.-K. Wang (Bytedance)]

Items 1 and 2 of this contribution belongs to this category.

TheThe first aspectaspect concerns the concept of “incomplete” random-access AUs and was initially deferred for offline study. It was further discussed on 23 April at 0840.

Decision (expression of existing intent): Specify that GDR AUs shall be complete – i.e., all of the layers in the CVS shall have a picture in the AU (as with IRAP AUs).

See section 6.1.3 for aspect #2.