JVET-N0754 BoG report on Coded Picture Regions [S. Deshpande (Sharp)]

The BoG met on 20 Mar 2019 from 14:30 to 20:45, on 21 Mar 2019 from 9:00 to 10:00 and 4:20 PM to 9:10 PM, and on 22 Mar 2019 from 9:00 to 13:30 and 14:45 to 16:00. The status was then discussed in the high-level syntax / systems relation meeting of Saturday 23 March 0900-1230 (chaired by GJS and Youngkwon Lim).

Approaches for independently coded picture regions can be categorized as follows:

1. Motion-constrained tile sets (MCTSs)
   • Encoding-constrained, such that encoding is constrained in a manner that a rectangular set of tile groups can be extracted and repositioned without affecting their decoding or decoding of other tile groups.
   • Independent rectangular tile groups (IRTGs): Encoding and decoding are modified to treat IRTG boundaries like picture boundaries. Implications on IRTG boundaries are handled in low-level decoding process.

Encoding-constrained MCTSs and IRTG are similar except how they treat boundaries. The term MCTS is used below for this category

1. Sub-pictures (non-layer approach): Sub-pictures are treated like pictures in the low-level decoding process. Sub-pictures of the same timestamp are spatially arranged onto an output picture.
2. Independent sub-picture layers (in the sense of the way HEVC uses the term “layer”, what HEVC calls independent non-base layers): Decoded pictures of the same timestamp and of different layers (each potentially having different SPS, potentially different CTU sizes and tool enabling/disabling) could be either
   • Composed into the same output picture or
   • Output separately for use by the system for different purposes (e.g., view scalability, texture versus depth, etc.)

Decision (bug fixes): The BoG recommended, and Track B agreed, to the following:

• Adopt an inference rule for single_tile_per_tile_group_flag syntax element when not present, from JVET-N0124 Proposal 1 (also reported in JVET-N0071).
• Adopt an inference rule for num_tiles_in_tile_group_minus1 syntax element in tile group header from JVET-N0288 Proposal 1.

In the discussion, it was noted that there are some differences between our tile groups and HEVC MCTSs – e.g., we have a one-to-one mapping of tile groups to NAL units, whereas an HEVC MCTS has no particular relationship to NAL units.

For “sub-pictures”, it was asked whether there would be anything fundamentally different about these from the decoder perspective, versus just a grouping of tile groups (possibly as IRTGs). It was said that there is no intent to have a significant difference.

The above aspects were reviewed in the high-level syntax / systems relation meeting of Saturday 23 March 0900-1230. Further meetings of the BoG were planned to occur.

Further discussion was held Tuesday 26 March at 1500 (chaired by GJS).

The BoG had also met on 24 Mar 2019 from 10:40 to 12:30, on 25 March 2019 from 18:00 to 22:15, and on 26 March 2019 from 19:30 to 22:15.

The BoG recommended the following:

• The BoG recommended Track/Plenary discussion on whether WPP in a form affecting the decoding process is included in the VVC design.
  • It was asked how the experience had been with WPP in HEVC. It was commented that it had been designed well in terms of avoiding significant difficulty of implementation. It was commented that WPP had been used substantially in the market, especially for software encoders and decoders, while others said it had not been necessary and that there were other ways to achieve parallelism. It was asked whether its use in the bitstream should have been mandatory. It was said that it was appropriate that it was not mandatory to be enabled. Decision (parallelism): Include WPP. Whether to prohibit some combination with tiles or not is for further study.
  • If the WPP feature is included in VVC, the BoG recommends adoption of JVET-N0150 (option 1). HEVC has a 2-CTU lag, and this proposal shows 0.2% penalty relative to a 2-CTU lag and 1% penalty relative to not using WPP. If the CTU size is smaller, the penalty relative to a 2-CTU lag was reported to be about 0.6%. It would be possible to have both the 1-CTU lag and 2-CTU lag schemes in the design, but others suggested just having the 1-CTU lag scheme. With the larger CTU size of VVC relative to HEVC, it seems important to minimize the lag in units of CTUs. Decision: Adopt (option 1).
  • If the WPP feature is included in VVC, the BoG recommends adoption of JVET-N0149. Text is provided in v3. It is proposed to allow WPP to be switched on and off on a per-tile-group basis for rectangular tile groups, motivated by the desire for extraction and merging of content from different encoders. For raster scan tile groups, the flag would not be sent. It was commented that this issue is not unique to WPP – every coding option that can be enabled or disabled at level higher than the tile group level in a way that affects the decoding process. Further study is needed to determine how we will approach this issue in general. It was commented that it is not uncommon for there to need to be some high degree of control over what an encoder will do when operating in a mix/extract system environment. For WPP, the proponent suggested that an encoder might want to enable WPP on large tiles and disable it on smaller ones.
• BoG recommended to adopt from JVET-N0852 the change in constraints as described as a more accurate representation of the Marrakech meeting notes. This had become obsolete based on other decisions of this meeting.
• The BoG recommended adoption of a layer concept without inter-layer referencing in JVET-N0278-v3 with following changes: Constraint flags are: optional in VPS and retain it as mandatory in SPS as in current VVC WD
• Decision (functionality): Adopt. Software (not necessarily merged into the main branch) is to be made available by J. Boyce within one month).
• VPS presence optional; the target layer is the first layer in the bitstream if the VPS is not present.
• Aspects agreed for further study:
• Parameter set sharing across layers.
• HRD parameters signalling and HRD model
• BoG recommended adoption of JVET-N0857 as an implementation of a “design in principle” which was agreed in Track B as noted in the main meeting notes.
  • It was noted that software for this had not been implemented, although the software actually supports slices, which are a superset of the functionality.
  • Regarding uniform spacing, it should work the same way for tiles widths and heights and for brick heights. The issue raised in the plenary was further discussed and agreed not to be so relevant, and typically a system would want widths and heights to be exactly equal as much as possible. It was confirmed that the approach of JVET-N0498 would be applied. Each tile will have a flag for whether it is decomposed into bricks and if so, there would then be a flag for uniformity, and if so, one height (minus 1) would be sent. If not , the number of bricks (minus 2) and then N-1 heights (minus 1) would be sent. Decision (more typical as default): Adopted as described.
  • Decision (functionality): Adopt. Software (not necessarily merged into the main branch) is to be made available by K. Suehring within one month).

Further BoG meeting was held Tuesday evening in room K.

The outcome was then discussed in plenary Wed ~1000 (GJS & JRO).

• BoG recommended to adopt proposal 2 of JVET-N0124, to modify to signal the bottom right tile index as a delta from top left tile index. Decision: Adopted.
• BoG recommended adoption of option 1 of JVET-N0047 with modifications which allow signalling reference picture lists for IDR pictures controlled by a SPS level flag. It was said that this is the same as BLA, just using the IDR NUT value. Decision: Adopted.

Tiles and tile groups (13)

JVET-N0754 BoG report on Coded Picture Regions [S. Deshpande (Sharp)]

See the notes in section 6.18.3.1.