JVET-AM0176 EE2-related: ECM NNLF evaluation using AhG11 trained models [T. Poirier, F. Galpin (InterDigital)]

This contribution reports the performance of the models trained and cross checked in AhG11 using NNVC software using the interface similar to the one used in NNVC software. The results on top of ECM 17.0 are as follow:

VLOP3:

• AI configuration {-0.79%,-1.81%, -1.72%} bdrate gains for Y,U,V with {100.1%, 223.0%} Encoding and decoding time
• RA configuration {-1.16%, -1.51%, -1.25%} bdrate gains for Y,U,V with {101.1%,205.1%} Encoding and decoding time

VLOP3 with TDO:

• RA configuration {-0.94%, -1.40%, -1.16%} bdrate gains for Y,U,V with {101.1%,150.9%} Encoding and decoding time

VLOP3 with TDO alternative lambdas:

• RA configuration {-X.XX%, -X.XX%, -X.XX%} bdrate gains for Y,U,V with {XXX.X%, XXX.X%} Encoding and decoding time

LOP5:

• AI configuration {-1.86%, -5.81%, -5.85%} bdrate gains for Y,U,V with {100.3%, 404.2%} Encoding and decoding time
• RA configuration {-2.54%, -5.13%, -4.70%} bdrate gains for Y,U,V with {101.2%, 337.6%} Encoding and decoding time

LOP5 with TDO:

• RA configuration {-2.35%, -5.05%, -4.70%} bdrate gains for Y,U,V with {101.2%, 241.5%} Encoding and decoding time

HOP5:

• AI configuration {-4.88%, -0.65%, -1.94%} bdrate gains for Y,U,V with {115.7%, 13928.2%} Encoding and decoding time
• RA configuration {-7.01%, -1.49%, -2.65%} bdrate gains for Y,U,V with {114.0%, 9890.9%} Encoding and decoding time

Contribution for information, no specific action required. The results indicate that for LOP without using TDO, the coding gain could be increased compared to EE2-4.5, but also the decoder runtime increase further. It is noted that the method used in TDO/alternative are different from those in 4.4/4.5.

It is noted that with some TDO settings, LOP may have a better tradeoff compression vs. decoder run time than VLOP.

It is commented that from the perspective of decoder manufacturers, TDO would not have impact on worst case decoder runtime and implementation cost.