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

JCTVC-M0256 Suggested fixes for HEVC Range Extensions draft text [D. Flynn, J. Sole, T. Suzuki] [miss]

Doc was still missing at the time of closing the meeting.

It was suggested that this contribution did not need to be presented in detail, as the proposed editorial improvement had been superseded by other decisions.

JCTVC-M0298 Bandwidth reduction for range extension [V. Seregin, X. Wang, J. Sole, M. Karczewicz (Qualcomm)]

(Initial discussion chaired by D. Flynn.)

HEVC Range Extensions 4:4:4 potential applications include consumer applications, so the complexity increase due to the 4:4:4 processing has to be taken into account. The scope of this contribution is the bandwidth of 4:4:4 motion estimation and compensation, which is around 37% greater than the 4:2:0 case for 4×2 memory block. Specifically, in this contribution, restriction of bi-directional prediction for 8×8 chroma components is studied, while luma component still can have bi-prediction. This modification reportedly increases bandwidth by 9%, as opposed to 37%, relative to 4:2:0. The impact on coding performance is reportedly about 0.2% and 0.4% luma BD rate loss for Main-tier RA and LB, and High-tier RA and LB respectively in for both YUV422 and YUV444. The impact on chroma was reportedly about 0.3–1.1%, depending on the test case.

The contribution proposes two methods, the first restricts 8x8 Bi-pred for all components, including luma, and the second proposes to implement the restriction just for motion compensation of the chroma components. It is suggested that this only is applied to 4:2:2 and 4:4:4 chroma formats.

It was suggested that there could conceivably be two 4:4:4 profiles, one targeting consumer applications and the other at higher operating points. Should this then depend upon the profile? Yes, as in some cases or application areas it may not be an issue.

Support was expressed for the proposal, remarking that there are application areas which would benefit from this.

Adoption of the second method (8x8 Bi-pred restriction for chroma) was initially recommended.

In further discussion, a participant remarked that the introduction of differences between the handling of luma and chroma in the 4:4:4 case is undesirable.

However, other differences were noted to exist – in regard to:

• 
  The motion compensation interpolation filters, for which shorter filters are applied to the "chroma" components, and
  
• 
  In-loop filtering, for which simplified filtering is applied to chroma.
  

The modification was noted to cause a conversion from a bipred prediction applied to luma to a mono-prediction applied to chroma, as the chroma MVs are derived from the luma MVs.

The loss from completely disabling bipred for 8x8 was substantially larger.

It was asked whether the test sequences include any fading sequences, which are likely to benefit from biprediction.

Further study was encouraged. It is acknowledged that memory bandwidth is a major concern and should be analyzed.