Of itu-t sg16 wp3 and iso/iec jtc1/SC29/WG11

JCTVC-O0199 SHVC skip pictures [J. Samuelsson, J. Enhorn (Ericsson)]

This contribution expresses support for the skip_picture_flag from contribution JCTVC-N0209 with the addition of handling special cases. In the method proposed for skip pictures in JCTVC-N0209, motion compensation is performed by applying the re-sampled motion field from the closest reference layer to the corresponding temporal reference pictures in the current layer’s DPB. This contribution agrees with the assessed advantage in using motion information from another layer while keeping fidelity (details) from the current layer. In addition to what is proposed in JCTVC-N0209, this contribution proposed the following:

• 
  Use re-sampled sample values from the inter-layer reference picture when the picture used for inter prediction in the inter-layer reference picture has no corresponding picture in the enhancement layer or when motion vector prediction is not used from the inter-layer reference picture.
  
• 
  Use inter prediction with zero motion from the first picture in L0 (and L1 in case of a B slice) when no inter-layer reference picture is available.
  
• 
  Bitstream restriction to prohibit the use of skip picture when neither inter-layer reference pictures nor temporal reference pictures are available.
  
• 
  Allow deblocking filter to be applied to skip pictures.
  

The proponent noted that the proposal might not fulfil the HLS-only concept, but presented some of the rationale and use case discussion for information.

The contributor discussed adaptive resolution conversion (ARC) and its desire to have a "diagonal" referencing across both time and resolution, whereas our draft requires only cross-layer or cross-temporal referencing.

The number of bits saved by having a special skip indicator instead of using CTU-level skipping was not estimated.

It was noted that we have a flag single_layer_for_non_irap_flag (in the VPS) for the ARC case.

(It was remarked that, historically, the skipping functionality had also been proposed for coding efficiency purposes, but this did not seem like a strong case for it.)

It was remarked that there might be a case for having an indicator of a mode of operation in which the bitstream is only allowed to have a non-skipped picture at one layer at a time.

It was remarked that enabling a diagonal prediction might be a cleaner approach for the ARC case. (A diagonal referencing was proposed in L0188.)

It was noted that the diagonal case could have better delay characteristics, assuming the encoder would not want to send many skips repeatedly in case it wants to switch up in the subsequent picture.

However, another participant remarked that enabling diagonal referencing would have a serious impact on various aspects of the design as it exists.

Regarding the indicator described above (a non-skipped picture at one layer at a time), the contributor suggested to change the semantics of single_layer_for_non_irap_flag. The indicator could also be in an SEI message. It was remarked that it might be desirable to put the indicator in the same place as single_layer_for_non_irap_flag, and that perhaps both could go into VPS VUI. It was suggested to put these as two flags in VPS VUI.

It was remarked that O0251 is related, as it discusses ARC.

It was suggested that the coding efficiency impact of requiring the encoder to provide the low-level bits for skipping an EL picture should be studied. There was some information provide previously in I0394 for frame doubling of 4k video – reporting 568 bits with skipped CTUs versus 96 bits, or with tiles 1464 bits versus 96 bits. For 1080p HD, 280 versus 96 and 1056 versus 96. It seemed like it might be better to not have the high-level alternative, if that is the rationale and it is within an encoder.

Further discussion Thu 31st (GJS). The text described above was provided in a -v3 version of the contribution and reviewed. An alternative expression in the –v4 verion using more explicit constraining of the syntax by the decoder was also discussed. The group preferred the more explicit signalling. Decision: Adopt -v4.