13.1.1.1.1.1.1.1.1JCTVC-C293 TE11: Summary report for TE11 on motion vector coding [J. Jung (Orange Labs)]
This document summarized the activities in Tool Experiment TE11 on motion vector coding. A total of 9 companies or universities had registered to the TE: Sony, Sungkyunkwan University, Samsung, NTT DOCOMO, LGE, HHI, Sharps Labs of America, ETRI / Kyung Hee University and Orange Labs. A few emails were exchanged mainly to coordinate the cross-check activity and exchange the modified versions of the TMuC software. Also it can be noticed that for some topics, either the contribution was cancelled by the proponent (lack of time running the experiment, or insufficient results), or the cross-check was not fully completed (all configurations, or sequences).
Several experiments on motion vector coding had been performed for TE11 with various configurations of the software.
For the Inter mode, it was reported that the MV Competition scheme provides significant gains.
For the Skip, Direct, Merge modes, both Skip + Direct (with MV Competition) and Merge were reported to bring significant gains. The following options were reported to seem possible for the TM:
-
Keep Merge only at CU and PU level
-
Keep Skip + Direct only at CU level
-
Keep Merge at PU level, Skip and Direct at CU level
When MV Competition is concerned (for either Inter, Skip and Direct modes) it was reported that:
-
Having a temporal predictor brings gains
-
Having the ability to signal the set of predictors (including the ordering of the predictors) brings gains
At this stage, experiments combining IMVP and MV Competition were reported to have failed to achieve noticeable improvements.
13.1.1.1.1.1.1.1.2JCTVC-C289 TE11: Report on experiment 3.1.a: Disabling MV competition [J. Jung, G. Clare, S. Pateux (Orange Labs)] (missing IPR statement prior)
The following remarks and observations were recorded:
-
Removing the MV competition on Inter, and Skip/Direct modes reportedly provides an average coding loss of 2.6% to 4.1%. (by tendency higher for RA and LC).
-
Decoder runtime reportedly increases by MV competition (not measured on the full set, only QVGA class D) 0.1-2.4%; with the highest increase for the lowest bit rate.
Conclusion: Keep MV competition for inter and skip, considering its compression gain.
13.1.1.1.1.1.1.1.3JCTVC-C132 TE11: Cross-verification result of Orange Labs motion vector competition proposal (3.1a) by SKKU [J. Yang, K. Won, B. Jeon (SKKU)]
This cross-check report confirmed the results of JCTVC-C289.
13.1.1.1.1.1.1.1.4JCTVC-C290 TE11: Report on experiment 3.2.b: MV-Competition on Inter, Skip and Direct modes [J. Jung, G. Clare, S. Pateux (Orange Labs)] (missing IPR statement prior)
The TMuC reference applies MV-Competition for Inter modes, and Merge. Skip and Direct modes are disabled. This experiment evaluated a configuration with MV-Competition for Inter, Skip and Direct modes, with Merge disabled.
The document reported that the scheme with MV-Competition enabled on Inter, Skip and Direct mode provides on average better results than the current TMuC reference.
Average BR gain was reported as 1.5%, gain on LC was higher – around 2.5-3%.
13.1.1.1.1.1.1.1.5JCTVC-C043 TE11: Simulation results of merge/skip (3.2b and 3.2d) [Y. Suzuki, F. Bossen (NTT DoCoMo)]
The following remarks and observations were recorded:
-
Confirmed the results of JCTVC-C290 for the HE case (LC not done).
-
Reported increased encoder run time by 10%, with a marginal decoder time impact.
-
Experiment 3.2d : Switching off skip, direct, CU merge increases BR by roughly 5%, no impact on encoding/decoding time (only done for HE case).
13.1.1.1.1.1.1.1.6JCTVC-C292 TE11: Verification of experiment 3.2.d by NTT DOCOMO [J. Jung, G. Clare, S. Pateux (Orange Labs)] (missing prior, available first day)
The following remarks and observations were recorded:
-
Confirmed the results of JCTVC-C043 3.2d HE cases.
-
Additional results on LC were: RA: Skip 8% / Merge 6%, LD: Skip 10% / Merge 6%
-
For LC, no results on encoder/decoder run times were provided.
13.1.1.1.1.1.1.1.7JCTVC-C199 TE11: Report on experiment 3.2.c: Check skip and merge together [I.-K. Kim , T. Lee , W.-J. Han (Samsung)]
In this document, the coding efficiency of a block merging process (MRG) and Skip/Direct mode of TMuC software was presented. The bit rate impact of Skip/Direct mode was reported as -0.7% (random access), -2.6% (random access low complexity), 0.1% (low delay) and -2.9% (low delay low complexity) compared to MRG which is turned on by default in the TMuC. This contribution also presented a combination of MRG and Skip/Direct where Skip/Direct was enabled on top of MRG. The reported bit rate impact was -1.2% (random access condition), -2.4% (random access low complexity condition), -0.5% (low delay condition) and -3.3% (low delay low complexity condition).
The following remarks and observations were recorded:
-
Compared to configuration skip+direct on CU merge off, 0.5% BR gain on average if both enabled. No significant increase in encoder run time.
-
Skip+Direct vs. CU based merging seems not always to be identical in effect (as was assumed before).
13.1.1.1.1.1.1.1.8JCTVC-C191 TE11: Cross-check result of merge/skip (3.2c) S. Jeong, J. Lee, H. Y. Kim, S.-C. Lim (ETRI), K. Kim, H. Lee, G. Park (KHU)
This cross-check report confirmed the results of JCTVC-C199. The decoder run time was only marginally increased, encoder run time increased around 10+%.
It was remarked that the high gains reported in the LC case may be caused by deficiencies of LCEC vs. PIPE, and some of this may be resolved in "LCEC phase II".
13.1.1.1.1.1.1.1.9JCTVC-C166 TE11: Study on motion vector coding (experiment 3.3.a & 3.3.c) [K. Sato (Sony)]
This contribution reported results of two experiments as follows:
-
Experiment -1: Comparison of with & without Temporal Predictor
-
Experiment -2: Code Number Swapping of Motion Predictors
The following remarks and observations were recorded:
-
Experiments were performed only for 50 frames.
-
Average numbers were missing; for experiment 2 the gain seems to be around 0 overall.
13.1.1.1.1.1.1.1.10JCTVC-C116 TE11: Cross verification of Sony proposal on motion vector coding [Y.-J. Jeon, B.-M. Jeon (LG Electronics)]
This cross verification contribution confirmed the results of JCTVC-C166 on experiment 2 (with a reported gain of approximately 0); no cross-check was provided on experiment 1 in the contribution.
13.1.1.1.1.1.1.1.11JCTVC-C291 TE11: Report on experiment 3.3.b: "temporally oriented" set of predictors for MV-Competition [J. Jung, G. Clare, S. Pateux (Orange Labs)] (missing IPR statement prior)
This contribution reported that by changing the order of predictors (temporal first), additional gain is possible – in particular for static background scenes. The contribution included a suggestion to signal the set of predictors in MV competition (potentially different for skip as well) at the slice level. No results on that were available yet.
It was reported that this would come with practically no cost at the decoder, but the encoder would need to do some extra optimization work.
Further work was encouraged
13.1.1.1.1.1.1.1.12JCTVC-C308 Cross verification of MV Coding Proposal by JCTVC-C293 [K. Sato (Sony)] (late reg.) (verbally presented in TE11, uploaded after the meeting ended)
This cross-check contribution was registered very late. When brought up for discussion, it had not yet been uploaded and the contributor verbally reported that the work was not yet complete, but that he results achieved so far had confirmed JCTVC-C291. The contribution was eventually uploaded significantly after the meeting had ended.
Dostları ilə paylaş: |