JVET-AE0057 MTT split modes early termination [W. Ahmad, P. Wennersten, K. Andersson (Ericsson)]

This contribution presents an early termination method that is claimed to reduce encoding runtime for ECM. The cost of coding the CU without splitting the CU at all is compared to a threshold. If the cost is higher than the threshold, MTT split will not be considered, which is claimed to reduce encoding runtime. The proposal is implemented on top of ECM 9.0 software and compared to ECM-9.0, the average BDR loss and runtimes reported in this contribution are as follows:

Test1 (Proposed condition is applied to all QPs)

Y 0.03%, U 0.06%, V 0.11%, EncTime 97.0%, DecTime 99.9% for the RA configuration,

Y 0.05%, U -0.06%, V -0.20%, EncTime 96.8%, DecTime 100.1% for the LDB configuration,

Test2 (Proposed condition is applied to QP 22 only)

Y 0.00%, U 0.00%, V 0.02%, EncTime 98.8%, DecTime 99.9% for the RA configuration. Overall EncT & DecT for QP22 are 95.5% , 99.7% respectively

Y -0.01%, U -0.02%, V 0.02%, EncTime 98.6%, DecTime 100.1% for the LDB configuration. Overall EncT & DecT for QP22 are 95.5% , 100.4% respectively.

For the worst-case RA encoding time, it is claimed that the proposed scheme reduces the runtime by 3% for the slowest chunk in Parkrunning3 QP22. For the worst-case LDB encoding time, it is claimed that the proposed scheme reduces the runtime by 5% for BQTerrace QP22.

It is proposed to include the Test2 in the ECM for RA and LDB camera captured contents.

It was asked if the formula (including a term (QP-22)) was tested for QPs below 22.? No, but it is believed that the threshold by tendency must be higher in lower QPs, which would be the case.

It was asked if the threshold should be bit depth dependent? Was not considered so far.

Test 1 is not QP dependent, and not class dependent.

Test 2 is QP dependent, and not class dependent.

concern was expressed by one expert that this introduces a new QP dependent encoder optimization, of which some aspects are not yet fully understood.

For AI, there would be no effect, as MTT starts at 32x32.

Generally, the tradeoff of test 1 (encoder runtime decrease vs. bit rate increase, which is consistent over classes) is asserted to be attractive.

Performance for screen content? For test 1 Class F -0.02, TGM +0.19% (which may be interpreted as very small loss, as gain/loss in TGM is usually much larger than in other classes)

Decision (SW/CTC): Adopt JVET-AE0057 Test 1. (CTC also for optional classes). Also for VTM with encoder config option, but not yet CTC for VTM, as results are not yet complete.

It was requested to conduct further study on behaviour for lower QP, and bit depth other than 10, to make an assessment about adoption for VTM CTC also for extended ranges of bit depth and QP.

Post-meeting note: A crosscheck on this contribution had originally been registered as JVET-AE0215, but was never delivered. According to a verbal report of the crosschecker, partial results were matching, and the implementation was straightforward.