Hacker News new | comments | show | ask | jobs | submit login

Second option is rather unlikely, given how many unknowns you would be solving against (geometry of the seen, geometry of the not seen, incident light sources, camera response curve etc) from comparatively very little data (a continuous signal that is sampled and heavily quantized to just 8 bits per channel).

Good point. We could reduce some of those unknowns by getting the original photograph (would have camera data in its metadata), and we could assume at least some of the light sources in the ceiling and through windows.

Yes you can limit the problem until we can get something useful out of the other end but it is very very hard. Here is one of the latest works on the topic: "Accidental pinhole and pinspeck cameras: revealing the scene outside the picture" (Torralba & Freeman).


Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | DMCA | Apply to YC | Contact