International organisation for standardisation organisation internationale de normalisation

18.4Motion vector coding

JCTVC-C208 Simulation results of motion partition, block merging and motion competition schemes [I.-K. Kim, T. Lee (Samsung)]

In the current TMuC design, there are two approaches for frame partitioning: prediction unit (PU) splitting to make smaller motion partitions and block merging (MRG) to make larger motion partitions. In this proposal, anchor results are obtained by running TMuC-0.81 using reference configuration with disabling all motion partitions, motion competition and MRG. By enabling MRG on top of the anchor, improved BD bit rate is reported to be about 2.0% (random access), 0.2% (random access LC), 2.4% (low delay) and 0.0% (low delay LC). But higher BD bit rate savings can reportedly be obtained by enabling motion partition: 4.2% (random access), 5.8% (random access LC), 4.1% (low delay) and 5.1% (low delay LC). By enabling motion competition, improved BD bit rate savings are reported to be about 4.3% (random access), 5.6% (random access LC), 3.2% (low delay) and 5.2% (low delay LC). The best performance can reportedly be obtained by combining motion partitions and motion vector competition which shows 7.2% (random access), 9.2% (random access LC), 6.2% (low delay) and 8.3% (low delay LC) BD bit rate savings.

The results of TE12 were reported to have been confirmed.

JCTVC-C257 On motion vector competition [Y. Su, A. Segall (SHARP)]

A method was proposed to improve the performance of motion vector competition in applications that may experience errors during transmission. The focus of the contribution was the dependency between parsing and co-located motion vectors that is incorporated into the motion vector competition design. This dependency results in coding gain; however, it also results in an inability to parse a bit-stream if a previous frame is lost. To overcome this problem, it was proposed to send the motion vector candidate as a flag, which signals if the selected predictor is the temporal collocated block, and subsequently send an index into the set of spatial predictors when the temporal collocated block is not selected. Remarks recorded in the consideration of the contribution were as follows:

• 
  This was presented near the end of the meeting, after difficulty coordinating a prior presentation time
  
• 
  The goal is reduction of the buffer size; the advantage of robustness under loss was not proven yet.
  
• 
  In the worst case the buffer is around 33% of a picture buffer in size.
  
• 
  This may require computation of a median of 16 values, which adds complexity.
  
• 
  A loss around 0.1% in bit rate was reported.
  
• 
  It was suggested to include further investigation of this in a CE
  

18.5B picture reference list redundancy

JCTVC-C278 Redundancy reduction in B-frame coding at temporal level zero [B. Li, J. Xu, G. J. Sullivan, F. Wu (Microsoft), H. Li (Univ. of Sci. &Tech. China)]

In the current TMuC, the usage called GPB is widely used to replace a P slice with a B slice using two reference frame lists that contain the same pictures. As list 1 is identical to list 0, there is redundancy comparing using only list 0 with using only list 1 as references. This contribution suggested not to use list 1 prediction in this kind of B slice, restricting the available predictors to using bi-prediction and list 0 prediction. Compared to TMuC0.7 anchors, such a change reportedly leads to 0.1% bit rate savings for random-access cases and 0.9% bit-saving for low-delay cases, on average.

Encoding time was also decreased compared to the anchor (around 10%). See further notes regarding the JCTVC-C285.

JCTVC-C285 Modified uni-directional inter prediction in generalized P and B pictures [W.-J. Chien, P. Chen, X. Wang, M. Karczewicz (Qualcomm)]

This contribution proposed a modification of uni-directional inter-prediction in generalized P and B pictures (GPB). The modification restricts the reference list selection for uni-directional inter-prediction. Simulation results reportedly showed that 0.88% BD rate reduction can be achieved. The same method was proposed in JCTVC-C278.

It was remarked that the encoding time decrease is no argument, as this is implementation specific

At first it was suggested for an AHG to study this further and check the two methods, work out syntax and provide stable cross-checked software and results, and report about implications on the MMCO and RPLR

Decision: After further study and discussion, it was agreed that this seems to be quite simple to implement without complications (as confirmed by SW coordinator), and the proposal was adopted.

18.6Quantization control

18.6.1.1.1.1.1.1.1JCTVC-C135 Flexible scaling of quantization parameter [D. Hoang (Zenverge)]

The current Test Model under Consideration (TMuC) employs a quantization parameter (QP) scaling that is borrowed from the AVC standard. In AVC, the quantization step size increases by approximately 12.25% with each increment of QP, so that the quantization step size doubles when QP is incremented by 6. For the purpose of rate control, the contributor asserted that this 12.25% increment may be too coarse for certain applications, such as low-delay coding. This contribution proposed a specification that allows the granularity of the quantization parameter (QP) to be varied at the slice and picture levels. Backward-compatibility with the AVC approach is maintained at one of the granularity settings. In addition to varying the granularity, this contribution also proposed a variable QP offset that can be specified at the slice and picture levels.

It was remarked that there have been prior efforts to address the coarse quantization step size increment of AVC. In JCTVC-A114, the quantization step size was proposed to double every 16 increments of QP.

In AVC, the QP can vary from 0 to 51, inclusive (for 8 bit per sample coding). A full transition of QP from 0 to 51 reportedly translates to a 362-fold increase in the quantization step size. The contributor said that, in practice, only a small range of QP values are used when coding a scene at a given bit rate. It was also asserted that at scene changes, QP can change due to a change in scene complexity and then typically varies only slightly within a scene.

It was remarked that the complexity of supporting the proposed inverse quantization process in the decoder may not be trivial. It was suggested that dynamic range increase may be needed to obtain the suitable precision for frequency-specific norm adjustment, and that necessary LUT sizes and perhaps some other memory increase may be needed for the decoder to support this.

The relationship between luma and chroma quantization was also discussed.

It was noted that the stability of the transform design may not be at a stage where such a proposal would be timely.

Further study was encouraged.