I can see how it may potentially work with fine crafted bloom filters - lossy, of course! I'm just not sure if training the set of filters would be worth the effort, because computing the filters might take a looooooooooooooooong time on that level of information complexity. I've been toying around with something like that over the years actually, had some models I've programmed (not for pictures though) - I gave up.
Sounds like the kind of thing that a company like Google could tackle. It would be especially good in the mobile world, where bandwidth is more limited.
Well funded organization could do it for sure. And Google would benefit from it, of course. But, from my limited research over the years I'm not convinced anymore that any further significant advancement could be done via continuous mathetmatics. It might be my personal bias from my experience, but I have a strong hunch on new approaches via discrete mathematics will be the key for any serious advances approaching shannon's limit / kolmogorov complexity (if those hold true, and they should as far as we know).
