Subject: Re: potential gamma bug present in OpenCV
Date: Monday 1st March 2010 12:33:51 UTC (over 6 years ago)
Hi, Well, I'm strongly against such automatic gamma handling. And I would not call the current implementation buggy. Many basic image processing functions in OpenCV can handle 16-bit, even 32-bit floating-point images, so we can accurately represent scenes with a very wide dynamic range using gamma=1. I do not mind if cvCvtColor (some of the conversions) will be extended to support gamma, optionally. cvShowImage, cvSaveImage, cvLoadImage can be extended too. But the other processing functions should process data as-is. If the function is called cvAdd, it should do C = A + B, not C = (((A/255)**(1/2.2) + (B/255)**(1/2.2))**2.2)*255. We just can not afford it, cpu-clock-wise and programming+debugging-time-wise. I myself never ever got into this problem with gamma; I've seen only one such case in my practice, on MacOSX, where the image codecs apply the gamma automatically - so we got different results in different OSes. After we switched to libjpeg on MacOSX, the problem was eliminated completely. Regards, Vadim On Mon, Mar 1, 2010 at 8:27 AM, Guy K. Kloss <[email protected]> wrote: > Hi, > > I'm presuming that OpenCV suffers from the same type of bug. I have NOT > verified it, but extrapolating from colour handling in other places within > OpenCV lets me assume that image scaling, smoothing, etc. do not take into > account any influence of a gamma coefficient. > > This should be something that can be fixed easily, by adding a default > gamma=2.2 parameter to some functions, that a user can override upon need. > It > would be better than using no gamma correction *at all*. > > One may just now also argue in one of two ways *against* this: > > * If all other software operates with this bug, why should we care to > introduce an incompatibility? > > * Undoing a gamma correction before the operation, and adding it > afterwards > again adds computational overhead that wasn't in there before (and makes > OpenCV slightly slower than the previous version). > > However, I think it would be a worthwhile undertaking to add this > correction > to the code base (in case the bug is in OpenCV as well), to be rather > correct > than just make the same mistake the likes of PhotoShop are making. > > Any comments? > > Guy > > ---------- Forwarded Message ---------- > > Subject: [Image-SIG] gamma bug present in PIL 1.1.7 > Date: Mon, 01 Mar 2010, 16:44:20 > From: Bill Janssen <[email protected]> > To: "[email protected]"