On Sun, 6 Oct 2019 16:40:37 +0200, =?UTF-8?Q?Aur=c3=a9lien_Pierre?= wrote: > argh. Tales of over-engineering…
I don't really disagree with you, just want to point out that getting it anywhere near correct (i. e. without a huge number of false positives and false negatives) is a difficult problem. > Just overlay the euclidean norm of the 2D laplacian on top of the > pictures (some cameras call that focus-peaking), and let the > photographer eyeball them. That will do for subjects at large aperture, > when the subject is supposed to pop out of the background. For small > apertures, the L2 norm will do a fair job. And it's a Saturday afternoon > job, hence a very realistic project given our current resources. That's fair, I just think that this kind of algorithm will likely select a lot of photos that are badly out of focus (because the focus locked on a much more expansive background) and miss ones where it's the relatively small subject that's in focus. > What you ask for is AI, it's a big project for a specialist, and it's > almost sure we will never make it work reliably. The drawback of AIs, > even when they work, is they fail inconsistently and need to be > double-checked anyway. > > So, better give users meaningful scopes and let them take their > responsibility, rather than rely on witchcraft that works only in > Nvidia's papers on carefully curated samples. Or maybe just implement focus peaking, as you say, but with a UI similar to the camera's UI (flashing regions that are in best focus). Then it's up to the user to select the best photos based on their knowledge of the desired subject. > Le 06/10/2019 à 16:18, Robert Krawitz a écrit : >> On Sun, 6 Oct 2019 15:02:39 +0200, =?UTF-8?Q?Aur=c3=a9lien_Pierre?= wrote: >>> That can be easily done by computing the L2 norm of the laplacian of the >>> pictures, or the L2 norm of the first level of wavelets decomposition >>> (which is used in the focus preview), and taking the maximum. >>> >>> As usual, it will be more work to wire the UI to the functionality than >>> writing the core image processing. >> Consider the case where the AF locks onto the background. This will >> likely result in a very large fraction of the image being in focus, >> but this will be exactly the wrong photo to select. >> >> Perhaps center-weighting, luminosity-weighting (if an assumption is >> made that the desired subject is usually brighter than the background, >> but not extremely light), skin tone recognition (with all of the >> attendant problems of what constitutes "skin tone"), and face >> recognition would have to feed into it. -- Robert Krawitz <r...@alum.mit.edu> *** MIT Engineers A Proud Tradition http://mitathletics.com *** Member of the League for Programming Freedom -- http://ProgFree.org Project lead for Gutenprint -- http://gimp-print.sourceforge.net "Linux doesn't dictate how I work, I dictate how Linux works." --Eric Crampton ___________________________________________________________________________ darktable developer mailing list to unsubscribe send a mail to darktable-dev+unsubscr...@lists.darktable.org
