WaveOne

[willvarfar]

I bumped into this on the Internet and, finding there was no mention of it here yet, thought it needed its own thread

http://www.wave.one/icml2017

"As of today, WaveOne image compression outperforms all commercial codecs and research approaches known to us on standard datasets where comparison is available. Furthermore, with access to a GPU our codec runs orders of magnitude faster than other recent ML-based solutions: for example, we typically encode or decode the Kodak dataset at over 100 images per second."

[necros]

too bad no price, no trial version

[nikkho]

too bad no price, no trial version

I guess that it is even not yet implemented.

[Jarek]

I have mentioned a few months ago that there is coming revolution in multimedia compression with GANs ( https://en.wikipedia.org/wiki/Genera...arial_networks ) ... and here it is.

their arxiv: https://arxiv.org/pdf/1705.05823.pdf

[Piotr Tarsa]

Isn't that going to distort images in unpredictable ways? I mean, assuming it relies on some pre-trained models, then the reconstructed details will depend on the dataset on which the program was trained on. So if the program was trained on dogs and we're compressing cats, then as the compression ratio goes higher we'll have dogs' features reconstructed in cats' images. Or I'm missing something and there are no pre-trained models in this compression scheme?

[wilon]

Can someone knowledgable here read and tell it's real or bullshit? There's a part like this in the paper:

Finally, PiedPiper has recently claimed to employ ML techniques in itsMiddle-Out algorithm (Judge et al., 2016), although their nature is shrouded in mystery

But all their other arguments seems very sensible and serious, I'm not yet clever enough to tell if this is a serious paper or a joke one. Can we summon Matt here?

[Bulat Ziganshin]

Isn't that going to distort images in unpredictable ways? I mean, assuming it relies on some pre-trained models, then the reconstructed details will depend on the dataset on which the program was trained on. So if the program was trained on dogs and we're compressing cats, then as the compression ratio goes higher we'll have dogs' features reconstructed in cats' images. Or I'm missing something and there are no pre-trained models in this compression scheme?

it's well-known fact that people trained on recognizing europeans' faces, are less successful in recognizing faces of asians and vice versa. so, why not? just select a proper training set if you are going to compress cats

[Jarek]

It is not a problem to do the training right - accordingly to types of images it will be actually used for - the arxiv says they use Yahoo Flickr Creative Commons 100 Million dataset.
It is used here by GAN's discriminator network to fit the features to the space of real-life images.

ps. the Silicon Valley reference is obviously a joke, but the rest seems legit ... this TV series has brought some (very skewed) spotlight to data compression field ... at least that everybody in this industry wants to be seen as the Pierd Piper ...

[Alexander Rhatushnyak]

http://www.wave.one/icml2017

"Here is how different image codecs compress an example 480x480 image to a file size of 2.3kB"

Guess their 2.3 kB image must be friends with a 10 mB compressed model file, and furthermore there should be hundreds of such models.