Abstract not available in document
JVET-C0101 Report of BoG on QTBT configuration setting [K. Choi]
This document provides a report on the study of QTBT configuration settings. The meeting was held in C2 at 9:00–10:00.
Mandates
- To further study detailed results of the EE
- To study possibilities of reducing the encoder complexity
- To define default settings of QTBT for CTC (e.g. max CTU size, separate or non-separate trees for luma and chroma in case of intra).
Mandates 1 & 2
Parameters regarding QTBT
- CTUSize: Basic unit structure of CBs by a QTBT structure
- MinQTLumaISlice: Quadtree leaf node size for luma in I slice
- MinQTChromaISlice: Quadtree leaf node size for chroma in I slice
- MinQTNonISlice: Quadtree leaf node size in B or P slice
- MaxBTDepthISliceL: Binary tree depth for luma in I slice
- MaxBTDepthISliceC: Binary tree depth for chroma in I slice
- MaxBTDepth: Binary tree depth in B or P slice
During the EE study of QTBT, three tests (i.e., Test 1, Test 2, and Test 3) were tested and some of the results were available. The configuration set and results are shown in the tables below.
- Anchor: JEM 2.0
- Test Sets: QTBT code in EE branch with Test 1, 2, and 3
Configuration sets of QTBT
Param. | Test1 | Test2 | Test3 | Test4 | Test5 |
CTUSize | 128 | 128 | 128 | 256 | 256 |
MinQTLumaISlice | 16 | 8 | 8 | 8 | 8 |
MinQTChromaISlice | 4 | 4 | 4 | 4 | 4 |
MinQTNonISlice | 16 | 8 | 8 | 8 | 8 |
MaxBTDepthISliceL | 4 | 2 | 3 | 2 | 3 |
MaxBTDepthISliceC | 0 | 2 | 3 | 2 | 3 |
MaxBTDepth | 4 | 2 | 3 | 2 | 3 |
MaxBTSizeISliceL | 32 | 32 | 32 | 32 | 32 |
MaxBTSizeISliceC | 16 | 16 | 16 | 16 | 16 |
MaxBTSizeBPSlice | 128 | 128 | 128 | 128 | 128 |
Available test results
Measure | Test 1 | Test 2 | Test 3 | |
AI | BD-rate (%) | −3.3%/−5.6%/−4.6% | −2.0%/−10.4%/−9.8% | |
Encoding time (%) | 561% | 251% | ||
Decoding time (%) | 110% | 110% | ||
RA | BD-rate (%) | −3.8%/−8.9%/−8.3% | −1.6%/−9.9%/−9.7% | |
Encoding time (%) | 2xx% | 137% | ||
Decoding time (%) | 1xx% | 108% | ||
LD | BD-rate (%) | −4.5%/−5.0%/−5.8% | −2.9%/−4.6%/−5% | −3.9%/−5.6%/−6.0% |
Encoding time (%) | 241% | 131% | 194% | |
Decoding time (%) | 106% | 109% | 106% | |
LDP | BD-rate (%) | −4.4%/−5.3%/−5.6% | −2.8%/−4.7%/−5.0% | −3.8%/−5.8%/−6.3% |
Encoding time (%) | 221% | 118% | 175% | |
Decoding time (%) | 118% | 118% | 118% |
Note that the better performance of test 2 in chroma, compared to test 1, is due to the increase of MaxBTDepthISliceC from 0 to 2. It might be interesting to modify test 2 with this parameter also set to 0 to bring the encoder runtime further down. (denoted as “Test 2a”) further below.
Test1 (the configuration of the EE) had been cross checked by two companies, and the results were matched with the result of proponent. Test 2 with AI had been cross checked, but the other scenarios had not been cross checked. The proponent provided additional test result based on Test 3 configuration, but the set has not been cross checked yet.
Two additional test sets were suggested to verify the performance of QTBT (i.e., Test 4 and 5). However, these would likely have even higher runtime than Test 2 and Test 3.
During the BoG meeting, people agreed that practically testing all test sets rapidly is impossible especially for Class A sequences (4k) because testing Class A requires more than 2 weeks.
It is suggested that the other test sets are evaluated at the next meeting by establishing an ad hoc group regarding QTBT encoder configuration.
After the BoG meeting, one company said that they will Test 3 during this meeting. Table 3 shows the volunteers to test configuration sets from Test 2 to 4.
Testers of QTBT during this meeting
Test 1 | Test 2 | Test 3 | |
Tester 1 | Samsung (Available) | Samsung | Qualcomm |
Tester 2 | Qualcomm (Available) |
Mandate 3
Default settings of QTBT for CTC were agreed to be decided based on the test results obtained during this meeting
Initial recommendations by the break-out group
- To test and verify Test 2 and Test 3 during this meeting.
- To define default settings of QTBT in an additional BoG meeting during this meeting
- To establish an ad hoc group to get the best configuration setting for QTBT
Initial results were presented in JVET Fri afternoon. It was confirmed that the BoG should further investigate the Test 2 and Test 3 cases and report results later. It was agreed that if it is unrealistic to get full results for class A during the meeting, results with partial sequences should be reported along with an analysis of how homogeneous the performance in Test 1 was over the entire sequence.
If possible, additional results should be reported for Test 2a.
The BoG met again on Monday afternoon and provided the following results and recommendations.
Configuration sets of QTBT
Param. | Test1 | Test2 | Test3 | Test4 | Test5 | Test6 |
CTUSize | 128 | 128 | 128 | 256 | 256 | 128 |
MinQTLumaISlice | 16 | 8 | 8 | 8 | 8 | 8 |
MinQTChromaISlice | 4 | 4 | 4 | 4 | 4 | 4 |
MinQTNonISlice | 16 | 8 | 8 | 8 | 8 | 8 |
MaxBTDepthISliceL | 4 | 2 | 3 | 2 | 3 | 2 |
MaxBTDepthISliceC | 0 | 2 | 3 | 2 | 3 | 0 |
MaxBTDepth | 4 | 2 | 3 | 2 | 3 | 2 |
MaxBTSizeISliceL | 32 | 32 | 32 | 32 | 32 | 32 |
MaxBTSizeISliceC | 16 | 16 | 16 | 16 | 16 | 16 |
MaxBTSizeBPSlice | 128 | 128 | 128 | 128 | 128 | 128 |
Available test results
Measure | Test1 | Test2 | Test3 | |
AI | BD-rate | −3.3%/−5.6%/−4.6% | −2.0%/−10.4%/−9.8% | |
Encoding time | 561% | 251% | ||
Decoding time | 110% | 110% | ||
RA | BD-rate | −3.8%/−8.9%/−8.3% | −1.6%/−9.9%/−9.7% | *−3.0%/−11.4%/−12.6% |
Encoding time | *259% | 137% | *206% | |
Decoding time | *105% | 108% | *107% | |
LD | BD-rate | −4.5%/−5.0%/−5.8% | −2.9%/−4.6%/−5% | −3.9%/−5.6%/−6.0% |
Encoding time | 241% | 131% | 194% | |
Decoding time | 106% | 109% | 106% | |
LDP | BD-rate | −4.4%/−5.3%/−5.6% | −2.8%/−4.7%/−5.0% | −3.8%/−5.8%/−6.3% |
Encoding time | 221% | 118% | 175% | |
Decoding time | 118% | 118% | 118% |
(*: Average result of Class B, C, and D)
Test 1 had been cross checked by two companies and the results were matched with the result of proponent. Test 2 with AI had been cross checked, but the other scenarios had not been cross checked. The proponent provided additional test results based on the Test 3 configuration, but the set had not been cross checked yet.
Two additional test sets were suggested to verify the performance of QTBT (i.e., Test 4 and 5).
Among 5 test sets, Test 2 is expected to show the lowest encoding time. The group agreed on testing Test 2 only during the meeting and see the results in additional BoG meeting.
It was suggested that the other test sets be evaluated at the next meeting after establishing an ad hoc group regarding QTBT encoder configuration to generate the results.
After BoG meeting, one company said that they could run Test 3 during this meeting. The table below shows the volunteers to test configuration sets from Test 2 to Test 6.
Testers of QTBT during this meeting
Test 1 | Test 2 | Test 3 | Test 6 | |
Tester 1 | Samsung (Available) | Samsung | Qualcomm | Ericsson |
Tester 2 | Qualcomm (Available) |
Two companies confirmed the test results of Test 2 and Test 3 which were provided by the proponent. The results were not fully covered by the test results of proponent, but the available results were matched with the proponent results. Additionally, one company tested the Test6 case voluntarily.
The analysis of complexity and coding performance was done in Tests 1, 2, 3 and 6. Test results are available in an Excel file attached to the BoG report. The following graph shows the trade-off between encoding time and coding performance in RA.
Complexity order:
- Test 1 > Test 3 > Test 2 > Test 6
Coding performance order:
- Test 1 > Test 3 > Test 2 > Test 6
Trade-off between performance and encoding time
- Test 3 was considered to show the best trade-off if luma & chroma results are calculated
Trafficflow in Test 2 shows 1.5% coding loss in RA.
The QTBT proponent said that the updated result of Test 1 in RA (Class A1) showed 256% of encoding time which is less than previous results.
Opinions were surveyed and discussed.
Opinion 1: Using Test 1 as the default setting:
- One expert thinks that coding performance of Test 1 is attractive.
- It was suggested to use Test 1 as default setting for JEM 3.0, because there are several months before the next meeting, and we can have some fast algorithm developed by the next meeting.
Opinion 2: Using Test 2 as default setting:
- The increase of encoding time seems to be fine.
- Using Test 2 or 3 is more desirable due to the encoding time.
Opinion 3: Using Test 3 as default setting:
- Some experts said that compromised solution would be Test 3.
- Some experts suggested to use Test 3 with reduced number of frames for Class A1 and A2 (RA and LD) for testing for the next meeting.
- It is suggested to use a modified intra period for test sequences.
The group consensus was to use the Test 3 configuration setting as the default setting for CTC of JEM 3.0.
Mandate4
It was confirmed that the current QTBT version in the software branch for the EE does not support adaptive QP in JEM 2.0.
Recommendation by BoG group
- To use the Test 3 configuration as default settings of QTBT for JEM 3.0
- To establish an ad hoc to get the best configuration setting and develop a fast method for QTBT
- To fix localized delta QP signalling in the current QTBT version in the software branch for the EE for JEM 3.0
From the follow-up discussion in JVET:
Some concern was expressed about encoder complexity, but no objection was made against the recommendation.
Decision: Establish “Test 3” configuration as CTC for JEM3.
The software coordinators are tasked to additionally report results on the “Test 1” case for JEM3.
Sharp (T. Ikai) volunteered to report results on “Test 2” additionally for JEM3.
These additional test points will be valuable to assess the achievements of the AHG.
Delta QP will be implemented by Mediatek as a separate branch 2 weeks after the release of JEM3. The method of signalling at the PPS and block level shall be identical with the current method in HEVC (where a quantization group indicates the granularity of QP sharing by blocks).