The default encoder is the HEVC reference encoder. That means a) encode speed isn't necessarily representative, b) all encoder-side trade-offs are 100% optimized for PSNR since that's what they used for evaluation during standardization.
I would trust this encoder to do well on PSNR, but encode speed and subjective quality should be well below what well-tuned HEVC can do.
The decoder is FFMPEG's. The BPG distrib doesn't include any of the platform-specific SIMD etc. optimized kernels but the speed should be reasonably representative. It's a fair comparison to say the stock IJG JPEG lib (which also doesn't use SIMD) for example.
According to some evaluation the HEVC committee did, the transform/coding unit size stuff has a pretty huge impact. See here - for the large videos ("Class A", 2560x1600 I believe), limiting CTU size to 16x16 (instead of their default 64x64) increases bit rate at same quality by 28.2%. So that's definitely the single largest thing at the image sizes you're looking at.
I would be surprised if the new prediction modes made more than 1-2% difference. They added more directional prediction modes, but these are fairly rarely used in H.264 to begin with.
Also note they went from a very approximate DCT ("hey, it's within 8% of the real thing!") to a much more accurate one. If my Bink 2 experiments are any indication that should be another 1% or so right there. But yeah, big blocks, man.
"12-08-14 - BPG"
3 Comments -
The default encoder is the HEVC reference encoder. That means a) encode speed isn't necessarily representative, b) all encoder-side trade-offs are 100% optimized for PSNR since that's what they used for evaluation during standardization.
I would trust this encoder to do well on PSNR, but encode speed and subjective quality should be well below what well-tuned HEVC can do.
The decoder is FFMPEG's. The BPG distrib doesn't include any of the platform-specific SIMD etc. optimized kernels but the speed should be reasonably representative. It's a fair comparison to say the stock IJG JPEG lib (which also doesn't use SIMD) for example.
December 8, 2014 at 4:32 PM
I am surprised how strong it is. I didn't think that H265 was that different from H264 for still frames.
My guesses for the main wins are the
1. block-size-pyramid & large block sizes
2. more intra prediction modes & better modes
3. ?
December 8, 2014 at 5:50 PM
According to some evaluation the HEVC committee did, the transform/coding unit size stuff has a pretty huge impact. See here - for the large videos ("Class A", 2560x1600 I believe), limiting CTU size to 16x16 (instead of their default 64x64) increases bit rate at same quality by 28.2%. So that's definitely the single largest thing at the image sizes you're looking at.
I would be surprised if the new prediction modes made more than 1-2% difference. They added more directional prediction modes, but these are fairly rarely used in H.264 to begin with.
Also note they went from a very approximate DCT ("hey, it's within 8% of the real thing!") to a much more accurate one. If my Bink 2 experiments are any indication that should be another 1% or so right there. But yeah, big blocks, man.
December 8, 2014 at 9:39 PM