Joint Collaborative Team on Video Coding (jct-vc) Contribution

5.12Quantization

JCTVC-B035 [X. Yu, D. He, E.-h. Yang (RIM)] Improved quantization for HEVC

This contribution proposed a quantization scheme for residual coding based on adaptive reconstruction levels, consisting of three steps:

• 
  First, a "hard decision" quantization is performed: quantization with a rounding offset being zero is conducted, and a reconstruction level is computed as the centroid for each quantization output.
  
• 
  Second, fixing the reconstruction level for each quantization output, an RDOQ algorithm is applied to re-calculate quantization outputs.
  
• 
  Third, given the quantization decision by RDOQ, the reconstruction levels are updated.
  

Simulation results reportedly show 0.2 to 0.5 dB gain, compared with RDOQ.

Non-uniform reconstruction was advocated. In this proposal, the encoder sends three values q1, q2, and q on a frame basis – separately for each block size (10 bits each). For i=0, the reconstruction is 0. For |i|=1, the reconstruction is sign(i)*q1. For |i|=2, the reconstruction is sign(i)*q2. For |i|>2, reconstruction is i*q.

This scheme was reportedly tested in the TMuC 0.2 context.

Intra MB usage was disabled for non-I frames, and the technique was not applied to I frames – this was due to the spatial prediction in I frames that causes the reconstruction to affect the prediction of the subsequent blocks.

A 2-15% bit rate improvement per sequence was reported for the non-I frames with this scheme (relative to intra-disabled reference). Only 8 frames were coded for each sequence.

It was remarked that a prior relevant contribution was JVT-P053.

It was remarked by a participant that both the distortion and bit rate change significantly relative to the reference. The number of bits was lower and the PSNR was lower when using the technique – which is somewhat like increasing the QP value.

The short sequence provides less opportunity for error propagation relative to the I frame. The gain seemed best for low-activity sequences – i.e., sequences that rely more on the I frame.

It was remarked that just having the encoder optimally choose a separate QP value for each transform block size might provide some gain.

A participant remarked that for rate control or perceptual reasons, an encoder would change the QP value within a picture – which would affect that ability to use this scheme.

A participant suggested adjusting lambda for a given QP.

A participant noted that the relationship between QP and lambda may be different in the TMuC than in prior designs.

The results seemed somewhat preliminary. There needs to be some way to deal with intra and spatially-adaptive QP selection. And it should be tested relative to using a larger QP in the reference to produce a more similar bit rate and PSNR operating point.

However the concept seems interesting and potentially promising in some form.

Further notes:

Design like rate-constrained Lloyd Max but also taking into account the necessary rate for side info for encoding the reconstruction table of a non-uniform quantizer by yet another lambda-times-rate term.

Proposal to signal only the two innermost reconstruction levels and the stepsize for the outer levels (each by 10 bits). This is done separately for each DCT size.

It is a two-pass encoding process: Encoding is done by uniform quantization first, computing centroids of two innermost reconstruction levels and the q value (for distance of the outer reconstruction levels). These are used for the non-uniform quantizer used in final quantization (RDOQ-like).

It is strange that in the reported results are lower in bitrate and in PSNR as compared to the TMuC results. Apparently, this corresponds to a larger QP value. This could have implications as it is similar to having a larger QP variation between I and P. (this explains relatively larger gain in low activity sequences such as Vidyo).

Only 8 frames were encoded per sequence, results reported for class C, D and E

The current method does not allow change of QP below slice level.

Intra disabled, currently only applied for P pictures, Rate gain reported counts only the P pictures.

Results are preliminary, but further investigation necessary on issues above to take any action.
JCTVC-B059 [J. Zheng (HiSilicon & Huawei)] Adaptive frequency weighting quantization in macroblock level

(refers to previous JCTVC-A111 and JCTVC-A028) Usage of parameterized frequency weighting models (instead of direct specification of quant matrices) at picture level – assign weighting parameters to frequency bands; those weighting factors are also shared by transforms of different block sizes. Switch quant matrix on or off at MB level. Syntax allows usage of different weighting modes (up to 7 excluding case of weighting off) that can be selected. Mode to be selected at MB levels can also be encoded dependent on modes of adjacent MBs, and on MB type.

Adapted once per frame.

Results reported from KTA experimentation using 2 different adaptive quant are around 3% BR saving for class B and C, only around 0.5% for class D.

Not compared against AQMS from KTA (which gives PSNR losses versus flat quant)

Not clear yet how it would be implemented in TMuC

Concern raised about restricted flexibility due to parameterized weighting

Not clear how weigthing parameters at picture level are determined. Orally it is said that it depends on number and relevance of coefficients found in the respective band; depending on the subjective importance of the band, the weight is either increased or decreased compared to the default (which is determined from a standard visual model)

Unclear how frequency weigthing can increase PSNR (that uses flat weighting of quadratic errors) – may this be due to the unequal quantization of picture types that is happening in the local adaptation?

The following aspects were recommend for further study:

• 
  Better explain how the model parameters are derived.
  
• 
  Propose how to combine this into TMuC.
  
• 
  Justify the necessity of up to 7 models.
  
• 
  Analyse how the effective quantization step size is adapted in various frame types.