International organisation for standardisation organisation internationale de normalisation

14.5TE12 Interpolation filters

14.5.1.1.1.1.1.1.1JCTVC-C050 TE12.2: Interpolation filter SIFO and DIF tested against MOMS by Fraunhofer HHI [B. Bross (Fraunhofer HHI)]

This document reported the results of testing the 12-tap SIFO and 6-tap DIF interpolation filters against 6-tap MOMS filtering within the scope of tool experiment 12 evaluating TMuC Tools. In a high efficiency configuration scenario, MOMS was tested against the SIFO default. Bit rate savings are highly sequence dependent. Thus, average BD-rates of -4.6% are shown as well as BD-rates of 12% when comparing MOMS to the SIFO reference. The average encoder runtime of MOMS is 62%-68% of the SIFO encoder runtime. On the other hand, the average decoder runtime of MOMS is 111% to 122% of the SIFO decoder runtime. In a low complexity scenario, MOMS was tested against the DIF default. MOMS performs better than DIF in terms of coding efficiency with average bit rate savings around 5% while encoder runtimes up to 139% and decoder runtimes up to 133% of the DIF runtimes have been measured.

The following remarks and observations were recorded:

• 
  Results show slight increase of BR (0.3%) for HE case, but reduction of encoding time by roughly 35%, increase of decoding time by roughly 15%
  
• 
  For LC (6-tap), rougly 6% BR decrease, increase encoding time by roughly 25%, decoding time by >30%
  
• 
  Note: It is reported that the results may be wrong due to a bug in the fixed-point implementation which is effective in class E (static background)
  
• 
  Sequence dependent, significant gain comes from BQSquare, BQTerrace, BlowingBubbles, PartyScene
  
• 
  Uses 12 bit precision in case of IBDI off, 16 bit otherwise.
  
• 
  Due to usage of IIR filter, also full-pel values are filtered.
  
• 
  Potentially this implies that 2 buffers are needed (as in the current TMuC implementation). Bringing it to one buffer would certainly affect the performance.
  

14.5.1.1.1.1.1.1.2JCTVC-C087 TE12.2: Experimental results of interpolation filter [T. Chujoh, K. Kanou, T. Yamakage (Toshiba)]

In this contribution, detailed results of interpolation filters were reported. This is one of reports in Tool Experiment 12: Evaluation of TMuC Tools. The interpolation filter of the high efficiency anchor is SIFO (Switched Interpolation Filter with Offset) and that of the low complexity anchor is DIF (Directional Interpolation Filter). In this experiment, some DCT-IFs (Discrete Cosine Transform – Interpolation Filters) are compared to anchor filters. The experimental results show that the coding gain of DCT-IF 12-tap compared to SIFO in the case of high efficiency anchor is average of 0.1% and the coding gain of DCT-IF 6-tap compared to DIF in the case of low complexity anchor is average of 3.6%. The encoding of DCT-IF 12-tap compared to SIFO in the case of high efficiency anchor is about 2/3 and the decoding time is reportedly decreased by 5%. The encoding and decoding times of DCT-IF 6-tap compared to DIF in the case of low complexity anchor were reportedly stable.

• 
  For HE cases, little gain of DCT-IF 12 (0.1% on average), reduction of encoding time by roughly 30%.
  
• 
  For LC cases, 1.7% gain on average for DCT-IF 6-tap, no significant increase in encoding time, for 8-tap filter 3.5% gain, increase of encoding time by 10%.
  
• 
  It was remarked that the performance seems to be highly variable over different sequences, e.g. roughly 14% for BQSquare in LD LC.
  

14.5.1.1.1.1.1.1.3JCTVC-C241 TE12: Crosscheck on interpolation filter: SIFO(DIF)/MOMS [R. Panchal, M. Karczewicz (Qualcomm)]

The following remarks and observations were recorded for this contribution:

• 
  Confirms results of JCTVC-C050, emphasize different treatment of chroma (higher precision filters) by MOMS (SIFO compared against 0.7)
  
• 
  Confirm sreport of JCTVC-C087, difference of TMuC 0.7 and 0.8 due to bug fix; 0.8 roughly reduces encoding time by 25%, BR by 0.15% for SIFO
  
• 
  Emphasizes that SIFO has additional advantage of illumination comp (not effective in current test set).
  

14.5.1.1.1.1.1.1.4JCTVC-C239 TE12: Report on interpolation filter: SIFO/DIF [R. Panchal, M. Karczewicz (Qualcomm)]

The following remarks and observations were recorded for this contribution:

• 
  Confirms results of Toshiba (C087)
  
• 
  Gain of SIFO12 vs. DIF6 in RA/LC 5.5%BR, 200%ET, 115% DT, LD/LC 5.9%, 176%/110%
  
• 
  Loss of DIF6 vs. SIFO in RA/HE 3.3%BR/59%ET/96%DT, LD/HE 3.9%BR/65%ET/97%DT
  

14.5.1.1.1.1.1.1.5JCTVC-C162 TE12.2: Results on MC interpolation filters [K. Kondo, T. Suzuki (Sony)]

This document reported the results of MC interpolation filters, especially SIFO/DIF as compared with DCT-IF for interpolation filtering in TE12. In the test, the coding performance and complexity were measured under common test conditions, which were defined in JCTVC-B300. The coding performance and complexity of MOMS were reported in addition to DCT-IF. These results were shown in the second part of the contribution. Detailed results were summarized in an attached Excel sheet.

The contribution confirmed the results of JCTVC-C087, with a slight difference due to different platforms.

Additional results:

• 
  for LC with 6-tap DCT-IF BR reduction 2.8% RA, 0.7% LD, no difference in ET & DT
  
• 
  for HE with 8-tap DCT-IF RA: BR increase 0.2% / 61% ET, LD: BR 1.1% / 65% ET
  
• 
  for LC with 4-tap DCT-IF 0.8% BR increase RA, 7.4% BR increase LD still ET roughly 100%
  

The contribution also confirmed the results of JCTVC-C050, with the following aspects noted:

• 
  Additional results on 4-tap MOMS for LC cases show BR red. 5.4%/6.1% BR red., 7%/6% ET inc. 10%/14% DT inc. for RA/LD cases compared to DIF anchors (note: MOMS with bug)
  
• 
  Confirms significant gain of MOMS in chroma due to use of longer filter (bilinear in reference cases)
  
• 
  Analysis of MOMS complexity: Due to usage of an IIR filter, the entire frame needs to be processed first. Memory complexity is significantly increased – higher memory access than conv. 8-tap.
  

14.5.1.1.1.1.1.1.6JCTVC-C197 TE12.2: Results for SIFO-6Tap with DIF tests by Nokia and Qualcomm [K. Ugur, J. Lainema (Nokia), R. Panchal, M. Karczewicz (Qualcomm)]

This contribution presented the Tool Experiment 12 results on Switched Interpolation Filter with Offset (SIFO) that switches separable and directional interpolation filters with 6 taps.

• 
  BR increase roughly 2% average for HE, BR decrease roughly 3% for LC
  
• 
  ET roughly 50% increase for LC, 10% decrease for HE
  
• 
  DT number not known, probably slight decrease in HE, slight increase in LC cases
  
• 
  Emphasized that switching between directional and 2D filters (at slice level) gives benefit (mostly for LC case)
  
• 
  Also reports advantage vs. DCT-IF 6-tap, particularly in LD LC gain is achieved over all classes.
  

14.5.1.1.1.1.1.1.7JCTVC-C201 TE12.2: Experimental results of SIFO (DIF) / DCT-IF [E. Alshina, J. Chen, N. Shlyakhov, A. Alshin, W.-J. Han (Samsung)]

This document reported the results of the chroma DCT-IF extension within the scope of tool experiment 12 evaluating TMuC Tools. For the test, DCT-IF 6 tap was used to replace the existing default chroma interpolation (bilinear filter for high efficiency and DIF for low complexity). The interpolation filter for luma remained as it was for both high efficiency and low delay configurations, which is SIFO for high efficiency and DIF for low complexity, respectively. Experiments results reportedly showed that, in the high efficiency test configuration, the average BD-rate reduction of Y, Cb, Cr components are 0.2%, 5.5% and 6.6%, respectively, for random access, and 0.5%, 2.5% and 2.9% for low delay. In the low complexity test configuration, the average BD-rate reduction of Y, Cb, Cr components are 0.4, 12.7 and 14.1%, respectively, for random access and 1.4, 12.4 and 15.3% for low delay. Encoding and decoding time was similar to that of anchor software, and deviations exist from measurement inaccuracies.

• 
  Encoding/decoding time increased slightly (2%Enc/7%Dec for LC case) – concern was expressed for this LC case.
  
• 
  BD BR saving in chroma may not be the best criterion for judgment, as the effect of better chroma interpolation is more about improving the PSNR of the chroma components, and the percentage of total bit rate for chroma is low.
  
• 
  The current default using bilinear interpolation seems to have a drift or rounding problem in chroma interpolation, and the gains reported may partially relate to the fact that this is resolved by higher-order interpolation.
  
• 
  The suggestion was made in the contribution to replace the 2-tap filter by DCT-IF 6-tap filter.
  

14.5.1.1.1.1.1.1.8JCTVC-C248 TE12.2: Cross verification on DCT-IF for chroma [T. Chujoh, K. Kanou, T. Yamakage (Toshiba)] (missing prior, available first day)

This contribution confirmed the results of JCTVC-C201.

JCTVC-C309 Cross verification of DCT-IF for chroma [T. Suzuki (Sony)] (late reg., uploaded after meeting had ended)

This late document was reported verbally during the meeting: It referred to JCTVC-C201, and was not finished yet when reviewed. However, the results reportedly confirmed the results appropriately for classes C and D. The actual document contribution was uploaded significantly after the meeting had ended.

14.5.1.1.1.1.1.1.9Overall summary of interpolation filter investigations:

• 
  It would be preferable to have the same chroma filter when doing an experiment where different luma interpolation filters are compared.
  
• 
  The issue of chroma interpolation should be further investigated and must be better understood.
  
• 
  MOMS: Good performance, but has open issues. Bug reported, claimed to potentially improve performance for class E. Bidirectional IIR filter requires doubling the memory, and resolving that may reduce the performance.
  
• 
  SIFO has highest objective gain typically, but higher encoder complexity.
  
• 
  For DCT-IF, DIF, SIFO in the various configurations (4..12tap) the group decided to provide a table about coding performance and enc./decoder complexity. Kemal Ugur was asked to coordinate a breakout to collect and organized the data to take as a basis for decision.