Back to Search Document details
16th Meeting: Geneva, October 2019 2019-10-07 18:17
Test results on removing sps_sbt_max_size_64_flag
Authors: Yin Zhao Huawei
Abstract
The sps_sbt_max_size_64_flag was designed to provide functionality similar to log2_transform_skip_max_size_minus2 for configuring encoder complexity. This contribution reports coding results of two tests of removing sps_sbt_max_size_64_flag with MaxSbtSize always being set equal to 64 in CTC. In test 1, an encoder only method is used to disallow SBT for CUs with one side greater than 32 for class C, D. and E, which tests the coding loss from encoder only method that achieves the functionality of sps_sbt_max_size_64_flag = 0. In test 2, the encoder only method is not used, which shows the increase of encoding time as well as the coding gain if always trying SBT for CUs with one side of 64 pixels. Test 2 is also the cross-check of JVET-P0391 table 4 (the code was kindly provided by the proponent of JVET-P0391), and the simulation results match with that in JVET-P0391.
Meeting Notes
JVET-P0983 Test results on removing sps_sbt_max_size_64_flag [Y. Zhao (Huawei)] [late]

The sps_sbt_max_size_64_flag was designed to provide functionality similar to log2_transform_skip_max_size_minus2 for configuring encoder complexity. This contribution reports coding results of two tests of removing sps_sbt_max_size_64_flag with MaxSbtSize always being set equal to 64 in CTC. In test 1, an encoder only method is used to disallow SBT for CUs with one side greater than 32 for class C, D. and E, which tests the coding loss from encoder only method that achieves the functionality of sps_sbt_max_size_64_flag = 0. In test 2, the encoder only method is not used, which shows the increase of encoding time as well as the coding gain if always trying SBT for CUs with one side of 64 pixels. Test 2 is also the cross-check of JVET-P0391 table 4 (the code was kindly provided by the proponent of JVET-P0391), and the simulation results match with that in JVET-P0391.

Test 1 results in 0.01% and -0.02% PSNR-Y-BD-rates in RA configuration for class C and D, respectively, and 0.01%, 0.02% and 0.13% PSNR-Y-BD-rates in LDB configuration for class C, D and E, respectively. The maximum loss is 0.29% for KristenAndSara sequence in LDB.

Test 2 results in -0.04% and -0.08% PSNR-Y-BD-rates in RA configuration for class C and D, respectively, and -0.03%, -0.01%, -0.11% PSNR-Y-BD-rates in LDB configuration for class C, D and E, respectively. The encoding time increase is around 2% and 3% for RA and LDB.

Both tests remove the sps_sbt_max_size_64_flag.

Test 1 is encoder only method, small loss in LB, but only for few sequences in class E

Test 2 sets the max size to 64, which does not have a loss, but increases the encoder run time for sequences with sizes smaller than HD.

There is no reason having a specific flag which helps an encoder for only few sequences. As test 2 would increase the encoder run time, and test 1 only affects 3 sequences, the test 1 method is preferable.

Decision: Adopt JVET-P0983 (removal of sps_sbt_max_size_64_flag).

Decision (CTC): Use test1 method, encoder enforces not using sbt for sequences below HD resolution.

References:
JVET-P0391
Cited:
JVET-P0991 X JVET-P2001 JVET-Q0002 JVET-Q0003 JVET-Q0041 JVET-Q0085 JVET-Q0097 JVET-Q0124 JVET-Q0125 JVET-Q0132 JVET-Q0148 JVET-Q0165 JVET-Q0168 JVET-Q0173 JVET-Q0175 JVET-Q0184 JVET-Q0185 JVET-Q0188 JVET-Q0199 JVET-Q0243 more than 20
ARCHIVE v2 :
JVET-P0983-v1.docx JVET-P0983-v2.docx JVET-P0983_SBT_ToolOffTest_AHG13.xlsm JVET-P0983_Test1_SBT32_encoderOnly.XLSM JVET-P0983_Test2_SBT64_always(xcheck).XLSM
Decisions
adopted
Adopt JVET-P0983 (removal of sps_sbt_max_size_64_flag)
Citation