Organisation internationale de normalisation

4.2Contributions

4.2.1.1.1.1.1.1.1JCTVC-D329 CE1: Samsung’s test for bi-directional optical flow [E. Alshina, A. Alshin, Woo-Jin Han (Samsung)]

This contribution related to CE1 Subtest 3. A coding gain of 1.9% for HE RA, and 4.8% for LC RA was reported. Roughly 12% and 100% increases in encoder and decoder runtimes for HE RA were reported. Roughly 11% and 150% increases in encoder and decoder complexity for LC RA were reported.

It was suggested that the concept presented in this contribution could be merged with concepts in other proposals to produce an improved design. It was not suggested that this proposal is sufficiently mature for adoption in its present form. Further study was encouraged (e.g., in an AHG).

4.2.1.1.1.1.1.1.2JCTVC-D121 CE1:Cross-verification report of Samsung's Proposal by JVC KENWOOD [Motoharu Ueda, Satoru Sakazume]

Cross-check for JCTVC-D329.

4.2.1.1.1.1.1.1.3JCTVC-D412 CE1: Cross-check result on Bi-directional optical flow of Samsung [Eunyong Son, Jiwook Jung, Sehoon Yea]

Cross-check for JCTVC-D329.

4.2.1.1.1.1.1.1.4JCTVC-D120 CE1:Refinement motion compensation using DMVD with merge extension [M. Ueda (JVC KENWOOD)]

This contribution related to CE1 Subtest 4. A coding gain of 1.6% for HE RA was reported. Roughly 50% and 40% increases in encoder and decoder runtimes for HE RA were reported. No gain was shown in the LD case. The decoder complexity increase is primarily due to ME operation in the decoder. Further investigation and potential harmonization with other concepts was suggested (e.g., in AHG).

4.2.1.1.1.1.1.1.5JCTVC-D340 CE1: Cross-verification of Kenwood’s experimental results of refinement motion compensation using DMVD (JCTVC-D120) by Samsung [E. Alshina, W.-J. Han, A. Alshin (Samsung)]

Cross-check for JCTVC-D120.

4.2.1.1.1.1.1.1.6JCTVC-D099 CE1: Report of implicit direct vector derivation [Yusuke Itani, Shun-ichi Sekiguchi (Mitsubishi)]

This contribution related to CE1 Subtest 1. A coding gain of 0.5% for HE RA was reported. A minor increase in encoder runtime, and 25% increase in decoder runtime were reported. No coding efficiency improvement was reported for the LD case. Further investigation and potential harmonization with other concepts was suggested (e.g., in AHG).

4.2.1.1.1.1.1.1.7JCTVC-D168 CE1: Cross-check of DMVD results from Mitsubishi (JCTVC-D099) [Y.-J. Chiu, L. Xu, W. Zhang, H. Jiang (Intel)]

Cross-check for JCTVC-D099.

4.2.1.1.1.1.1.1.8JCTVC-D295 CE1：Huawei report on TMDMVD and STDM in HM [Mingyuan Yang, Sixin Lin, Haoping Yu]

This contribution related to CE1 Subtests 1 and 2.

For Subtest 1, a coding gain of 2.0% for HE RA was reported. Roughly 70% and 6% increases in encoder and decoder runtimes were reported. If decoder template matching is added, an additional 0.3% coding gain was reportedly achieved, with higher complexity. It was suggested that some of the encoder complexity increase can be avoided. (It was noted that the complexity of the anchor can also be improved.)

For Subtest 2, a coding gain of 1.0% for HE LD was reported. Roughly 78% and 14% increases in encoder and decoder runtime were reported.

It was suggested that combination of this proposal with JCTVC-D167 may be feasible.

4.2.1.1.1.1.1.1.9JCTVC-D169 CE1: Cross-check of DMVD results from Huawei (JCTVC-D295) [Y.-J. Chiu, L. Xu, W. Zhang, H. Jiang (Intel)]

Cross-check for JCTVC-D295.

JCTVC-D295 has two tools:

• 
  Template matching based DMVD (TMDMVD) for GPB slices and
  
• 
  Spatial-Temporal Direct Mode (STDM) for non-GPB slices.
  

As such, TM-DMVD is tested in both LD and RA cases, but STDM is tested only in RA cases in JCTVC-D295.

Both STDM and TMDMVD-plus-STDM are cross-checked in JCTVC-D169.

Common conditions were followed. The results are complete.

The BD BR results in JCTVC-D169 and JCTVC-D295 match completely. The encoding/decoding times differ slightly (as typically expected).

Initially, JCTVC-D295 did not contain encoder runtime data. Upon checking with Huawei, a supplementary excel file was provided by Huawei and the encoder runtime was comparable to the cross-checked data.

Often, the cross-checker reported a lower increase of encoder/decoder runtime.

The cross-checker studied the software source code very carefully. They compiled and ran the program.

No irregularity was observed.

4.2.1.1.1.1.1.1.10JCTVC-D247 CE1: Cross Check with JCTVC-C501 (3.1.2) [Yue Yu, Krit Panusopone, Limin Wang] (missing prior, uploaded Thursday 20th, first day of meeting)

Partial cross-check for JCTVC-D295 (and JCTVC-C501 ‒ subset 2).

Only the Low delay (HE and LC) cases were cross-checked for JCTVC-D295 in JCTVC-D247.

In JCTVC-D247, the labelling of results had a mis-alignment problem, e.g. Class B was labeled as Class A.

With the mis-alignment problem compensated, the BD BR data of JCTVC-D247 reportedly matched the corresponding result of JCTVC-D295. The encoder/decoder runtimes also matched.

The software was not carefully studied by the contributor of JCTVC-D247 ‒ just compiled and run; however, since an additional cross-checker (JCTVC-D169) provided a more complete and very thorough analysis, this was not considered a serious problem.

No irregularity was observed.

4.2.1.1.1.1.1.1.11JCTVC-D167 CE1: Report of self derivation of motion estimation in TMuC 0.9 [Y.-J. Chiu, L. Xu, W. Zhang, H. Jiang (Intel)]

This contribution related to CE1 Subtest 1. A coding gain 1.7% for HE RA was reported. Roughly 23% and 4% increases in encoder and decoder runtimes were reported.

For a 2nd variation with motion search added, a coding gain of 2.9% for HE RA was reported. Roughly 34% and 19% increases in encoder and decoder runtimes were reported.

The relatively minor increase of decoder complexity was noted, although there was some interest expressed in being able to carefully study the software for the proposal to check how realistic this is (e.g., the degree to which software optimization is a part of this measurement).

4.2.1.1.1.1.1.1.12JCTVC-D098 Report of CE1 :Decoder-Side Motion Vector Derivation [Shun-ichi Sekiguchi, Yusuke Itani (Mitsubishi)] (missing prior, uploaded Tuesday 18th, before meeting)

Cross-check for JCTVC-D167 (JCTVC-C127).

JCTVC-D167 has 3 methods tested:

• 
  6-candidate,
  
• 
  6-candidate-plus-refinement,
  
• 
  9-candidate
  

The first two are cross-checked in JCTVC-D098. The third one is for info only, and not cross-checked.

In both JCTVC-D098 and JCTVC-D167, common conditions were followed. The results are complete.

The BD BR results in JCTVC-D098 and JCTVC-D167 match completely. The encoding/decoding time differs slightly, as expected in many cross-check situations, as different computers are used.

In general, the cross-checker in JCTVC-D098 showed lower increases in encoder/decoder runtime (e.g. 124% instead of 119%, 105% instead of 108%, 130% instead of 117%, 135% instead of 125%)

The cross-checker studied the software source code to some extent. They compiled the program themselves and ran the program.

No irregularity was observed.

4.2.1.1.1.1.1.1.13JCTVC-D446 CE1: Cross-verification report for the part of JCTVC-D167 proposed by Intel [Motoharu Ueda] (late registration Monday 24th after start of meeting, uploaded Monday 24th, fifth day of meeting)

Cross-check for part of JCTVC-D167 (successful cross-check).