Take a look at the examples in this new paper: https://arxiv.org/abs/1703.01467
They’re very interesting. I think there’s a lot of room to improve this technique, and you guys have all the info you need to tackle this problem now yourself. So maybe implementing and improving this paper and writing it up is another interesting project to try…
This is a fantastic paper! Thanks.
I liked and was equally amused by Figure 4. Since the picture is reconstructed from higher order learned features, at higher compression, the shoe type and shoe color itself changes. . The same is not that evident in the celeb database images though.
(in other words, if you have lower bandwidth, you get one shoe, and if you have higher bandwidth, you get another shoe.)
Right - I really think there’s a lot of room to improve on their results. Probably just a case of improving the loss function using the tricks we’ve been learning (i.e. other activation layers).