JVET-N0605 Comparative study of video coding approaches VVC, AV1, EVC versus HEVC [Y. Chen, E. François, F. Galpin, R. Jullian, M. Kerdranvat (Technicolor)] [late]

This was discussed Wed 27 1250 (GJS & JRO).

This contribution reports the results of a comparative study between video coding schemes including encoders for HEVC (ITU-T and ISO/IEC) and AV1 (Alliance for Open Media), or still in development such as VVC in the JVET group of IEC/ISO and ITU-T, and the new draft of EVC (Essential Video Coding) of MPEG. One-second clips were encoded.

A participant said that there were some problems in the configuration of AV1. The presenter acknowledged that it was not a completely fair comparison, generally in favour of AV1 in the testing.

The quantization applied for the I frames (key frames) and “alt-ref” frames was not controlled.

Two-pass encoding was used for AV1 (~13-15% benefit).

It was asked about how much extra gain may be from these issues in the AV1 configuration, and the author said perhaps about 23% (13-15% from two-pass, 3-4% from key frame QP control, 3-4% from adaptive GOP structure).

The methodology is somewhat based on the common test conditions (CTC) defined in JVET by using the same configuration parameters on all the sequences of classes A1, A2, B, C and D.

The performances according the average BD-rate PSNR metrics of VTM-4.0 (VVC after Marrakech meeting January 2019), AV1, and ETM 1.0 (EVC after Marrakech meeting) versus HM-16.18 was reported as 29.5%, 21%, and 15% gains for all the CTC clips. When considering only UHD and HD clips the measured gains were higher, 33.5%, 22.4% and 19.8% respectively. The run times versus HM-16.18 for encoding are respectively 7.4x, 6.7x, 3.6x and for decoding 1.2x, 2.5x, 1.4x.

In terms of decoding time, VTM performed relatively well in the test, about 20% slower than the HM.

The contributor remarked that high coding performance is important for success of the project.