Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Computational photography can be done on your computer. What you should ask yourself is why isn't there any widespread software for this sort of thing widely available? Because the sort of people who want overprocessed photos are happy with their phones.

People buy DSLRs for many reasons, none of which include unremovable sharpening filters, automatic replacement of faces with leafs or turning a bright orange sky blue. Or replacing the blurry 300px moon you just shot with a 3mp stock photo.



Does your computer know as much about the scene as your phone/camera at the time of shooting it? I really don’t think that a computer can do the same computational photography as a real-time device, at least not with current formats.

It’s probably the same as AOT vs JIT — the proper solution would be to move more data to the computer where we have faster hardware and less resource-constraints, which would be the equivalent of profile-guided optimizations in my analogy (things like movement, maybe even some basic description of the scene based on NN and previously seen images)


Well there's CHDK for Canon cameras. I had it installed on a point-and-shoot a few years back and really liked the features it gave.

https://chdk.fandom.com/wiki/CHDK


I did the same some 10 years ago only to discover I'm too lazy to do post-processing on my timelapse. You can definitely use it to record successions of stills and videos which I guess is part of what phones do. It would be better if there was native firmware support for some of that stuff though, and I imagine that some combination of firmware acquisition settings and accelerometer data would give you the same capabilities as smartphones. Perhaps maybe a secondary camera with a high framerate to get optical flow/camera pose data from? Dunno.

You can still do it manually, but give it 3 years and you'll be able to compose that photo from scratch with an off-the-shelf generative NN photoshop plugin anyway. But still I'd say the market for DSLRs will remain the same because their main function (in my opinion) is to capture an accurate representation of reality with predictable errors in pixel space and not in some latent one.


> automatic replacement of faces with leafs

this was confirmed to be fake news.


The leaf was in front of the face but the algorithm decided to blend it in that way.


For your computer to compete, it would need the full IMU and lidar stream, which is required for many of these computational tricks.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: