an open web media project
The latest news and updates from the WebM project

“Blueberry” VP8 Hardware Encoder IP Released

Thursday, May 12, 2011 | 10:04 AM

"Blueberry," the second release of the H1 VP8 hardware encoder, is now available through the WebM Project hardware page. Due to the short growing season and abundant light during the summer, Nordic blueberries are exceptionally sweet and rich with vitamins. The Blueberry encoder is not too bad either!

In Blueberry we focused primarily on improving the encoder for video calling use, as many of the semiconductor companies that have licensed the H1 encoder plan to use it in these use cases. Compared to Anthill, the average measured PSNR improvement was 0.82 dB, while SSIM figures were improved by 0.011. This is also shown in the following chart for 720p video call content, where Blueberry achieves the same quality as Anthill with 25% less bits!



In the next release, we plan to further improve the compression rate at the low bitrate range, as well as focus on new features such as two-pass encoding and visual optimization using segmentation maps. The third release is planned to be available at the end of Q2 2011.

The H1 IP has been licensed already to over twenty semiconductor companies, and we are looking forward to sharing the technology with new partners.

For licensing details about the H1, see our hardware page. For those interested in technical details, please keep reading here.

Aki Kuusela is Engineering Manager of the WebM Project hardware team in Oulu, Finland.


5 comments:

Igor said...

Please, accept constructive critics.

Why is PSNR still the main metrics?

It has been already discussed many times. The conclusion was the same: PSNR correlates bad with real visual quality.

Almost no improvement for SSIM.
It clearly shows that increasing PSNR numbers doesn´t mean quality improvements.

SSIM family metrics are much more realistic.

I hope PSNR metric won´t be used for visual optimization.

Artis said...

Why still no double blind testing anywhere in the community? Let people actually judge what looks good. No more declaring one objective metric to be superior to another, no more zooming in on stills. The audio community figured this out a while ago...

Aki Kuusela said...

Thanks for the comments.

Igor, the cost function of the hardware is now programmable, so we can optimize macroblock mode decisions for either PSNR or SSIM, we simply did the PSNR first. This is also the first hardware version to have segments supported (the feature in VP8 that allows macroblock level QP adjustment). So only now we can start to optimize the quantization based on psychovisual things like frequency masking.

Aki Kuusela said...

Artis, having people to do visual testing is obviously the best way to go in theory, but in practice it's impractical:
- hard to organize
- hard to repeat (subjects learn)
- takes too long (we need to run algorithm optimization runs a dozen time a day)
- single modifications to the encoder are too small to be visually distinguishable (consider that the 0.82 dB PSNR gain in Blueberry came from about 10 different improvements to the design)

Artis said...

Aki, it wasn't meant to be a critique of VP8 development practices. Cross codec/major version comparisons on which looks better as a video however are what should matter in the end and is what the discussions should be about, not which synthetic metric needs to be used to get there like Igor seems to imply.