International organisation for standardisation

6.11.2Quantization matrices

6.11.2.1.1.1.1.1.1JCTVC-F362 Proposal to support quantization matrix in HEVC [T. Suzuki, A. Tabatabai (Sony), M. Zhou, V. Sze (TI)

In this contribution, a quantization matrix that is simply extended AVC method is proposed to include in HM as a starting point of further investigation. Because of large block size transform and quantization, data size of quantization matrix can be large, and hence efficient encoding method to support quantization matrix is desired. The requirements and the current practices of quantization matrix are summarized in JCTVC-E056. During the discussion in Geneva, it was generally agreed that HEVC should have some form of quantization matrix capability. In this contribution, an extended AVC quantization matrix method is proposed to include in HM4. It consists of the followings.

• 
  Encoding of matrices in SPS and PPS
  
• 
  Support 4x4, 8x8, 16x16 and 32x32 matrices
  
• 
  DPCM + signed exp-Golomb code
  
• 
  AVC inverse quantization process
  

This method is implemented in HM 3.0 and the source code is attached to this contribution. The enhancement methods of the quantization matrix can be investigated based on this method. The quantization matrix should be disabled in common test condition if adopted.

Overhead is extremely high (in class D 5x video rate), as the proposal in the presented results uses random values and suggests to change these at every frame. This seems to be unreasonable.

There was agreement that quantization matrices should be included, and an “AVC like approach” should be used as a starting point (although this would not be used in the default setting). M. Karczewicz was asked to check the suggested solution and report back.

Default setting of these matrices must still be clarified -> the suggested solution was not considered reasonable.

One possibility would be to use the AVC matrices as starting point for 4x4 and 8x8, interpolated versions for the larger sizes? It should also be tested what the impact of the frequency of change is (only once per sequence, each xxth frame etc.). It may also be unrealistic that quantization matrices would be used with high QP values.

For further study in a CE.

6.11.2.1.1.1.1.1.2JCTVC-F085 Further study on compact representation of quantization matrices [M. Zhou, V. Sze (TI)]

Due to large block size transform and quantization used in the HEVC, carrying the quantization matrices in bitstream could lead to a significant overhead. In this document, a quantization matrix compression algorithm is proposed to achieve compact representation of quantization matrices for HEVC. Test results reportedly revealed that the algorithm provided up to 13-15x more compression when compared to the AVC quantization matrix compression method if all the compression tools are enabled. The contributor suggested to adopt the AVC quantization method into the test model, and start a CE on this subject to specify a compact quantization matrix representation format for efficient carriage of quantization matrices in HEVC.

6.11.2.1.1.1.1.1.3JCTVC-F475 Enhancement of quantization matrix coding for HEVC [J. Tanaka, Y. Morigami, T. Suzuki (Sony)]

Proposed requirements and reported current practices of quantization matrix are summarized in JCTVC-E056. In existing coding standards, e.g. MPEG2 and AVC, quantization matrix is introduced to improve subjective quality. Quantization matrix is widely used in consumer and professional video products, e.g. video cameras, Blu-Ray Disc, etc.

Data size of quantization matrix, especially for large CU, is asserted to be large, and efficient encoding method to support quantization matrix is advocated by the proponent.

If compression mode is DPCM mode only, as in AVC, it allegedly needs many bits. JCTVC-E073 was proposed in Geneva meeting. In the contribution, a coding method of the quantization matrix is proposed. Compared with AVC, the proposed method allegedly improves coding efficiency about 36.7% in average BD-rate. Coding efficiency is reportedly improved in all sequences and increased encoding time is small.

Conclusion: Start CE on quantization matrices as part of the quantization CE. It may not be necessary to include the AVC matrices in WD; an HM branch may be sufficient for the purpose of the CE.