search

jcelaya / hdrmerge

HDR exposure merging
http://jcelaya.github.io/hdrmerge/
Other
365 stars 78 forks source link

Hard transitions #6

Closed Beep6581 closed 6 years ago

Beep6581 commented 10 years ago

I'm really hoping that the issues I see are fixable, because there is enormous potential here!

So far on all shots I tried - Pentax K10D with manual or automatic lenses - I see very hard transitions where different source images were used. http://i.imgur.com/QSkBBxd.jpg http://i.imgur.com/xEIca0P.jpg

The colors in that house were very difficult as there was a sunlight with yellow glass camera-left, and the big window downstairs, plus other light sources and unlit areas (this is from a panorama, so it's only about 1/30th of the whole). http://i.imgur.com/O1wq05z.png HDRMerge with RawTherapee does a fantastic job if not for the hard transitions, and if invisible transitions were possible it would be a kickass solution!

See the large blobs on the ceiling and walls: http://i.imgur.com/THloiua.jpg http://i.imgur.com/j4tS6rF.jpg

I'm uploading a few bracket series here, some of the shots are from the Samyang manual lens without aperture info, so currently they won't work unless you use exiftool or make HDRMerge not rely on Exif :) The files should be up in 15 minutes. http://filebin.net/j5lo3p3s6m

jcelaya commented 10 years ago

I think this problem has something to do with #1 . There is no problem with the light color, because no color adjustment is done in HDRMerge. The white balance is passed untouched to the output DNG, and your camera selected the same white balance for all three exposures (which is common).

Beep6581 commented 10 years ago

Anything I can do to make a temporary workaround for my K10D shots? By the way, I tried with input PEF and DNG files, both resulted in hard transitions.

jcelaya commented 10 years ago

OK, try the last commit, it should work now. It seems your camera has some serious alinearities when recording values near the saturation level. This is common in other cameras too, so I implemented a more conservative method to calculate the relative exposure between images.

Beep6581 commented 10 years ago

Confirmed fixed, thank you! This is so great! :D

jannefoo commented 10 years ago

The problem still persist, at least with D800 DNG files. Not as severe anymore, but it's readily visible in images.

Beep6581 commented 10 years ago

jannefoo please show a screenshot of the problem and upload your source raw files, you can use www.filebin.net for that. Are you sure what you're experiencing is this bug, and not simply source image misalignment?

jannefoo commented 10 years ago

Here's a zip file with a merged DNG, source files and a screenshot from Raw Therapee showing the banding: https://dl.dropboxusercontent.com/u/168254285/web/hdr%20dng.7z

jcelaya commented 10 years ago

I see. It's more or less the same problem. I'll have a look at it next week.

Beep6581 commented 10 years ago

I ran into this bug again. :( http://i.imgur.com/K1ujb62.jpg http://filebin.net/wwobo2n8rk

Javier have there been any recent changes to this part of the code? If so, which version can I downgrade to so I can continue using it while you investigate?

jcelaya commented 10 years ago

I am working on it right now, but I have not uploaded anything yet. So, unfortunately, I don't think you can fix it by downgrading :(

Beep6581 commented 10 years ago

In the meanwhile I tried downgrading all the way back to february, and the same issues persists in every version. I also tried converting to DNG using Adobe DNG Converter 8.4, but no difference.

Funny that it uses the darkest image for those clouds when the +2 and even +4 can be used there.

Beep6581 commented 10 years ago

Hi! Have you had any luck?

jcelaya commented 10 years ago

Almost. For now, try not to use underexposed images.

Beep6581 commented 10 years ago

Will try that, thank you!

jcelaya commented 10 years ago

Well, I just pushed the solution to the develop branch. There were two problems here, actually:

Beep6581 commented 10 years ago

Sorry to bother you with this, but I ran into this issue again: http://i.imgur.com/65JXkGk.png http://i.imgur.com/gC4SZVq.jpg

Removing the brightest of the five bracketed images fixes the problem in both cases, but maybe there is something still wrong, something still fixable, so that spotting the problem in time and manually fixing it wouldn't be necessary?

I'll send you the link to the raws via email.

jcelaya commented 9 years ago

I commited some changes that improve the transition between images, but I still cannot get this last one correctly. Do you notice this problem often?

Beep6581 commented 9 years ago

If you consider one set of 5 images bracketed 2EV apart a "DNG", then I see this about once every 10 DNGs. As you know I shoot a lot of panoramas, and each panorama consists of about 6 DNGs.

I will test the commit on new panoramas and report back after a few days if I still find this problem. I generally have to test on new panos as I delete the individual PEFs which I've used for the old panos, to save space. Your previous suggestion of removing unnecessary bracketed photos when producing DNGs works quite well, though sometimes I don't notice the problem until it's too late.

Thank you :)

Beep6581 commented 9 years ago

Hey again!

I still run into this once in a while, but I learned to work around it, by removing the darkest images.

Today I ran into this issue again in a merge of three images. However in this case, the darkest image contains data in the highlights which is clipped in the middle image, so I need to use all three, yet I also get this hard transition in the top-left corner. You can get the images and the HDR DNG created with 0.5.0 with the problem here: http://filebin.net/b0w1sw908k The link will expire in a year.

jcelaya commented 9 years ago

I think this problem is similar to ghosting, but with light. Look at the upper left corner, for instance. Its brightness is similar to the rest of the ceiling in the middle image, but in the bright one all the ceiling is blown out but that corner. I'm still not sure how to fix this, though...

Beep6581 commented 6 years ago

The links have expired, the issues are mostly fixed, and the code changed significantly since 2014. Closing.