International organisation for standardisation organisation internationale de normalisation

14.3TE12 TMuC prediction units

14.3.1.1.1.1.1.1.1JCTVC-C052 TE12.1: Block merging (MRG) test (off vs. on) by Fraunhofer HHI [B. Bross (Fraunhofer HHI)]

This document reported the results of testing the block merging mode for CUs (MRG) within the scope of tool experiment 12 evaluating TMuC Tools. For the test, MRG was disabled and tested against the MRG-enabled default. In a high efficiency scenario, average bit rate savings up to 0.7% were reported when disabling MRG. In a low complexity scenario, disabling MRG yields average bit rate savings up to 2.9%. MRG disabled encoder runtimes range from 106% to 127% and decoder runtimes from 96% to 116% of the default MRG enabled runtimes have been measured.

Disabling CU based merging (reinvoking skip & direct) gives roughly 1% BR reduction overall.

14.3.1.1.1.1.1.1.2JCTVC-C109 TE12.1: Results for PU-based merging by Fraunhofer HHI [B. Bross (Fraunhofer HHI)]

This document reported the results of testing the block merging mode for PUs (MRG_PU) within the scope of tool experiment 12 evaluating TMuC Tools. For the test, MRG_PU was disabled and tested against the block merging mode for CUs (MRG) default. In a high efficiency scenario, average bit rate savings up to 1.9% were reported when disabling MRG. In a low complexity scenario, disabling MRG yielded average bit rate savings up to 2.4%.

The collection of experiment results was not yet complete. A "Bug fix" done in TMuC 0.8 on the PU based merging that was there before but was switched off by default (wrong RD decision at the encoder – ticket #95 and coding of PU based merging ticket #97 avoiding some unnecessary signaling). The results reportedly showed that PU based merging gives additional gain for the cases tested so far. Implicitly, this replaces CU based merging.

It was commented that proposals with similar goals are in TE11.

14.3.1.1.1.1.1.1.3JCTVC-C118 TE12.1 Cross-check report on block merging (MRG) on/off [K. Sugimoto, A. Minezawa, S. Sekiguchi (Mitsubishi Electric)]

This contribution confirmed the results of JCTVC-C052. Brief checking of parameter settings and software was performed.

14.3.1.1.1.1.1.1.4JCTVC-C231 TE12.1: Results for PU based merging [Y. Suzuki, T.K. Tan, F. Bossen (NTT DoCoMo)] (missing prior, available first day)

This contribution provided a cross check with more results than JCTVC-C109, with results reported as follows. For HE LD, 1.5% BR gain, 2% for HE RA. Also results were reported on case-by-case basis which show that using only skip/direct (as in JCTVC-C052) BQSquare, BQTerrace and Vidyo3 provide losses.

14.3.1.1.1.1.1.1.5JCTVC-C264 TE12: Evaluation of block merging (MRG) [W.-J. Chien, P. Chen, M. Karczewicz (Qualcomm)]

This contribution confirmed the results of JCTVC-C052. Brief checking of parameter settings and software was done.

It was remark that from the results, CU based merging does not seem to make sense. However, investigation of software related to CU based merging may not have been performed to the detailed extent that could definitely exclude the presence of bugs that would improve the performance if fixed.

14.4TE12 TMuC transform units

14.4.1.1.1.1.1.1.1JCTVC-C056 TE12: Evaluation of transform unit (TU) size [M. Zhou, V. Sze (TI)]

This document reported the evaluation results of the Transform Unit (TU) size for TE12, and provided a brief complexity analysis of transforms. Experimental results reportedly show that the average coding loss caused by disabling the 64x64 transform is around 0.1% to 0.6%; disabling both 32x32 and 64x64 transforms lead to an average coding loss of 1.1% to 4%. The results also reportedly show that disabling the 4x4 transform leads to huge coding loss. When both the 4x4 and 64x64 transforms are disabled, the average coding loss is increased to 3.9% to 8.3%, and disabling 4x4, 32x32 and 64x64 transforms (i.e. only 8x8 and 16x16 transforms are enabled) can lead to quality degradation of 5.2% to 10.6% on average. Due to the high complexity and limited coding efficiency gain of the 64x64 transform, it was recommended that the 64x64 transform be removed from reference configurations. The contribution also recommended that the quality issues of disabling the 4x4 transform be investigated in the future.

The following remarks and observations were recorded:

• 
  There is relatively small loss when 64x64 transform is disabled (for LD LC case which is worst of all 0.6% on average over all classes, 1.3% average for class E LD LC).
  
• 
  32x32 disabling causes larger loss (1 to 3.4% depending on case).
  
• 
  Disabling 4x4 causes larger loss compared to AVC.
  
• 
  Almost 4x gate count when 64x64 needs to be implemented compared to 32x32.
  
• 
  This could be dealt with as a profiling issue; the current test material seems not to benefit too much from using 64x64 (32x32 could be default in the TM, and further study of 64x64 could be performed in the future).
  

14.4.1.1.1.1.1.1.2JCTVC-C105 TE12: Huawei report on max transform unit size evaluation [J. Zhou, D. Wang, H. Yu (Huawei)]

This document summarized the Tool Experiment TE12 activities related to the evaluation of Max TU Size. This evaluation was performed by Huawei and TI.

• 
  Confirms the results by TI, with a small difference due to the use of Windows vs. Linux.
  
• 
  The contributor received the parameter files from TI and ran on newly compiled software.
  

14.4.1.1.1.1.1.1.3JCTVC-C200 TE12.1: Experimental results of transform unit quadtree/2-level test [T. Lee, J. Chen, W.-J. Han (Samsung)]

This contribution presented the experimental results of TE12 Residual quadtree transform (RQT), which was to explore the performance of RQT in the Test Model under Consideration (TMuC). Experimental results of RQT off compared to TMuC showed 1.2%, 0.5% and 0.2% bit rate savings under intra only, random access and low delay conditions, respectively, in high efficiency settings. It shows 0.6%, -0.2% and 0.8% bit rate savings under the same conditions, respectively, in low complexity settings. This contribution also analyzed the performance and complexity of the current TMuC transform unit design by testing several variants of it: no transform unit across partition boundary and no transform unit splitting inside a partition.

The following remarks and observations were recorded:

• 
  Turning RQT off gives loss in luma PSNR but also has less encoder complexity in the intra case (where the BR increase is 1.2%), in inter cases there is a smaller BR increase but no complexity advantage by turning RQT off.
  
• 
  Restricting TU to maximum of PU size (disallowing TU across PU partitions gives loss of around 1%.
  
• 
  Restricting TU not becoming smaller than TU size gives loss as well.
  
• 
  Conclusion: RQT on would only make sense provided that that the suggestion below would compensate for the complexity.
  
• 
  Intra prediction mode flag could also be used
  
• 
  Gain of TU across PU partition boundaries may be due to more efficient CBP
  
• 
  It could be that gain also comes mainly from asymmetric partitioning
  
• 
  A new suggestion (related to TE9_C198) was to limit the number of TU quadtree levels by using signaling.
  
• 
  JCTVC-C276 is also related as well.
  

These issues are further discussed elsewhere in this report.

14.4.1.1.1.1.1.1.4JCTVC-C110 TE12.4: Transform unit quadtree tested against 2-level by Fraunhofer HHI [B. Bross (Fraunhofer HHI)]

This document reported the results of testing the transform unit quadtree (RQT) against the 2-level representation within the scope of tool experiment 12 evaluating TMuC Tools. The 2-level TU representation was tested against the RQT default. BD-rates from -0.2% up to 0.7% were reported. The encoder runtimes of the 2-level representation vary from 31% of the RQT encoder runtimes for the intra only coding constraint set to 102% of the RQT encoder runtimes for the random access coding constraint sets.

This confirms the results of JCTVC-C200.