Adversarial nets allow that, by having random inputs. Somewhere, the neural net needs to decide "this car is going to be blue" and then be consistent with that. it might not know if the car is blue or red, so half the pixels would randomly be red, and the other have blue. But then sampling from it would be incorrect. Like a mixture of gaussians or something. You could have the net predict an entire distribution for each pixel. I wonder if there exists a cost function that directly promotes sampling from the underlying distribution without needing the adversarial approach. That's why the "Adversarial discriminator" should work. The optimal solution is simply drawing from the underlying distribution, instead of relying on a deterministic "best estimate": an outside observer won't be able to distinguish your generated samples from the true distribution. The ML estimate will pick the red car, which of course is unrealistic. ![]() So you might instead try a maximum likelihood estimate but this too has problems: imagine every car is a sightly different shade of blue (none are quite the same, maybe the manufacturing is unreliable), except 1 in a million cars are red, but the red is very consistent. The first thing you think of is giving the least-squares estimate (the average), but MMSE exhibits the problem shown. Let me try to confront the fundamental problem directly.Įven if you're given a perfect probability distribution over the space of images the solution wasn't obvious for me, mostly because we're used to thinking of a "best estimate". Your adversarial proposal is very interesting. Try a red filter and watch the blue sky become black!) (Even more of a tangent, one of the joys of B&W photography is that you can outright lie about colors and the photo still works. I haven't done any detailed research into this, but I'm not 100% convinced that collecting luminance through red, green, and blue filters can capture all the data that panchromatic B&W film captures. More subjectively, I think most digital pictures converted to B&W look kind of dull, whereas actual film looks very exciting to me. Lightroom, at least, did not let me make a matching image.) (Which should not be necessary, if you want to block blue and green, you should be able to do that in software. I could never make the RGB data from the digital camera look like my original negative, even when shooting through the same red filter. ![]() Most interesting was a case where I took a picture of the NYC subway diagram with a red filter. I've done a little bit of experimentation with this, comparing real B&W film to various software's conversions to B&W.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |