International organisation for standardisation organisation internationale de normalisation

14.6TE12 TMuC motion representation

14.6.1.1.1.1.1.1.1JCTVC-C030 TE12: Report on AMP evaluation [E. Francois (Technicolor), L. Guillo (INRIA), A. Ichigaya (NHK), H. Yu (Huawei)] (missing prior, available first day)

The following remarks and observations were recorded for this contribution:

• 
  Turning AMP off results in 1-2% BR increase for the 4 inter configurations, no large variations over the classes (slightly more gain in class E makes the LD cases a bit better).
  
• 
  Slight deviations of results were observed on Windows vs. Linux platforms.
  
• 
  Encoding time was reduced by roughly 33% by turning off AMP.
  
• 
  The relationship with other methods investigated was in TE3. Block partitioning appears useful, but not necessarily by AMP.
  

14.6.1.1.1.1.1.1.2JCTVC-C104 TE12: Huawei & HiSilicon report on AMP evaluation [X. Zheng (HiSilicon), J. Zhou, H. Yu (Huawei)]

This report confirmed the results of JCTVC-C030.

14.6.1.1.1.1.1.1.3JCTVC-C149 TE12.2: Report on asymmetric motion prediction unit (AMP) on/off [A. Ichigaya, K. Iguchi (NHK)]

This report confirmed the results of JCTVC-C030.

14.6.1.1.1.1.1.1.4JCTVC-C286 TE12.2: Technicolor & INRIA report on AMP evaluation [E. Francois (Technicolor), L. Guillo (INRIA)]

The report confirmed results of JCTVC-C030.

14.6.1.1.1.1.1.1.5JCTVC-C119 TE12.2 report on MV prediction AMVP/IMVP [K. Sugimoto, A. Minezawa, S. Sekiguchi (Mitsubishi Electric)]

The following remarks and observations were recorded for this contribution:

• 
  Switching IMVP on instead of AMVP produces BR increase (<0.5% on average, 1% for class A), and encoding time also increases slightly.
  
• 
  IMVP seems not to give value.
  
• 
  Further investigation of AMVP was performed in TE11.
  

14.6.1.1.1.1.1.1.6JCTVC-C163 TE12.2: Results on MV prediction [K. Sato, T. Suzuki (Sony)]

This contribution confirmed the results of JCTVC-C119, with identical results at the bitstream level (both run under Linux).

14.6.1.1.1.1.1.1.7JCTVC-C265 TE12: Evaluation of adaptive motion vector resolution (AMVRES) [W.-J. Chien, P. Chen, M. Karczewicz (Qualcomm)]

The following remarks and observations were recorded for this contribution:

• 
  Result were provided only for HE cases
  
• 
  TE12 result: the BR increase by turning AMVRES off is roughly 0.5%, the effect on encoding time was roughly 10%.
  
• 
  The new proposal of "unified AMVRES" that adapts to prediction direction results in – 1.4% difference on low delay, 0.6% for RA.
  
• 
  Additional results on GPB=1+ AMVRES=0 / GPB=0 + AMVRES=1 shows 4.7% / 3.5% decrease in BR and encoding time increase 215% / 112% compared to LD HE anchor.
  

14.6.1.1.1.1.1.1.8JCTVC-C182 TE12.2: Cross-check of adaptive motion vector resolution (AMVRES) technology in the TMuC software [A. Segall (Sharp)] (missing prior, available first day)

This contribution confirmed the results of JCTVC-C265.

14.7TE12 TMuC partition-based illumination compensation

14.7.1.1.1.1.1.1.1JCTVC-C041 TE12: Performance of partition based illumination compensation (PBIC) [A. Fujibayashi, S. Kanumuri, F. Bossen, T. K. Tan (NTT DoCoMo)]

This contribution was a TE12 contribution presenting the performance of Partition Based Illumination Compensation (PBIC) tool integrated into the TMuC software. In addition to the test conditions of TE12, the contribution compared PBIC to two other tools, namely SIFO and GRefMode, present in the TMuC software that can handle illumination changes. This comparison was also done when a linear fade was applied to the test sequences.

The results reportedly showed that PBIC has no impact on coding efficiency for the regular test sequences and a large gain (11-20% depending on the config) for the test sequences with linear fade. PBIC reportedly has a minimal impact (2-5% increase) on encoder/decoder run times. The authors said that this comparison indicated that PBIC is the top performing tool in terms of coding efficiency and has the least increase in encoder run time.

The following remarks and observations were recorded for this contribution:

• 
  For the common conditions test sequences, PBIC gives no gain or even produces a small loss (1%) in low complexity modes (could be due to the entropy coding).
  
• 
  For the same test sequences (first 2 seconds) with artificial fade (1 s fade out/in each), PBIC was better in terms BR reduction (12% on average) than SIFO (5%, TMuC) and GRefMode (10%, from A124), and less complex.
  
• 
  For more natural sequences with illumination variations (Nuts, Shuttlestart, Flower), a 2.6% benefit was reported on average.
  
• 
  It was remarked that cross fades may be more relevant than fade-in/-out
  

The PBIC proposal seems to be somewhat better than other potential methods of illumination compensation, but shows no gains on the current test set, and it is probably not a high priority to include dedicated tools for illumination variations.

14.7.1.1.1.1.1.1.2JCTVC-C232 TE12.7: Verification of JCTVC-C041 (partition based illumination compensation - PBIC) [N. Sprljan, S. Paschalakis, P. Wu (Mitsubishi Electric)]

This contribution confirmed the results of JCTVC-C041 on the common conditions set, and with some of the artificial fades, although there were some small deviations due to rounding errors. It was reported that the author had investigated the source code.

14.7.1.1.1.1.1.1.3JCTVC-C246 TE12.8: Cross-verification of partition-based illumination compensation (PBIC) [P. Onno (Canon)]

This contribution confirmed the results of JCTVC-C041 on the common conditions set.

14.8TE12 TMuC filtering

14.8.1.1.1.1.1.1.1JCTVC-C235 TE12: Evaluation of adaptive in-loop filter [I. S. Chong, M. Karczewicz (Qualcomm)]

The effect of the adaptive in-loop filter on BD measures was reported as follows:

• 
  High complexity
  

• 
  Intra : 2.2% (max 4.2% for class A)
  
• 
  Random access :3.7% (max 5.8% for class B)
  
• 
  Low delay: 5.2% (max 8.5% for class E)
  

• 
  Low complexity
  

• 
  Intra :-2.2%
  
• 
  Random access:-4.8%
  
• 
  Low delay: -4.3%
  

Decoding time increased by around 15% (10-20 depending on cases).

The report on encoding time (no significant difference) seems to contradict with the X-check, probably caused the the cluster environment which may not deliver reasonable numbers.

14.8.1.1.1.1.1.1.2JCTVC-C129 TE12: Cross verification for adaptive loop filter (ALF) (on vs. off) [Y.-J. Chiu, L. Xu, W. Zhang, H. Jiang (Intel)]

This contribution confirmed the results of JCTVC-C235 in terms of BR and DT; Encoding time seems also not to be consistent (increase by 30% switching ALF off?) due to usage of a computing cluster.

14.8.1.1.1.1.1.1.3JCTVC-C153 TE12: Report on deblocking filter [P. Chen, M. Karczewicz (Qualcomm)]

The following remarks and observations were recorded for this contribution:

• 
  Marginal changes in BD rate were observed.
  
• 
  The deblocking filter does not have significant impact on PSNR
  

14.8.1.1.1.1.1.1.4JCTVC-C148 TE12: Crosscheck on deblocking filter in TMuC0.7 by MediaTek [Y.-L. Chang, Y.-W. Huang, S. Lei (MediaTek)]

This contribution confirmed the results of JCTVC-C153 in BD rate.

The encoding/decoding time was roughly 3% lower when deblocking was switched off.

14.8.1.1.1.1.1.1.5JCTVC-C214 TE12.5: Results for adaptive loop filter using prediction and residual [M. Narroschke (Panasonic)]

This contribution was a part of the TE12: It presented the results of the Adaptive loop filter using the prediction and residual as filter inputs. An average Y-BD-rate gain was reported of 1.4% for Intra High Efficiency, 1.3% for Random Access High Efficiency, and 0.8% for Low Delay High Efficiency compared to the anchors. It was proposed to include it in the TM.

The test results were not complete yet due to recent bug fix work (the results were more complete in the cross check).

A gain of roughly 1% average (higher on intra) compared to ALF was reported.

Encoder complexity was roughly 25-30% higher than with ALF.

Decoder complexity was also increased by roughly 10%.

14.8.1.1.1.1.1.1.6JCTVC-C230 TE12: Cross check result of Panasonic's 3-input-ALF [I. S. Chong, M. Karczewicz (Qualcomm)]

This contribution confirmed the results of JCTVC-C214.

In comparison against ALF, the BD BR improvement was 1.2% for Intra Low Complexity, 1.1% for Random Access Low Complexity, and 0.9% for Low Delay Low Complexity

14.8.1.1.1.1.1.1.7JCTVC-C086 Experimental results of ALF on low complexity [T. Chujoh, T. Yamakage (Toshiba)]

The BR reduction in LC configurations was:

• 
  RA 4.83%/2.77%/7.56% for ALF alone / DCT-IF 6-tap alone / combination
  
• 
  LD 4.27%/0.67%/7.11% for ALF alone / DCT-IF 6-tap alone / combination
  

The suggestion of the authors was to adopt ALF in the test model.

It was remarked that the relationship of ALF and interpolation filters still needs to be better understood

JCTVC-C212 Cross-verification of JCTVC-C086: Experimental results of ALF on low complexity [W.-J. Han (Samsung)]

This contribution confirmed the BR reduction results reported in other contributions for this feature.

14.8.1.1.1.1.1.1.8Overall conclusions on filtering:

• 
  ALF results were consistent, with average rate gain slightly less than 5%, decoder runtime increased around 15%.
  
• 
  Encoder runtime reports were not fully consistent due to usage of clusters.
  
• 
  The relation between interpolation and loop filtering needs to be better understood.
  
• 
  Comparison against post filter (TE 10) should be considered.