> you did it, and what the target was, would be very helpful.
First, I'm a geologist, not a photographer, and while I may be close to expert at specialized aspects of photogrammetry, I am by no means expert at image processing or programming, though I have been tinkering extensively with both for more than 20 years. I know there are some great experts using darktable who could probably do much better than me, and I would dearly love any guidance anyone is willing to provide.
I did 91 different conversions from ARW or DNG to JPG or TIFF using a set of 120 highly overlapping aerial photos taken with a Sony A6R, and 30 conversions of a set of 1000 highly overlapping photos with the same camera. The goal was to maximize automated feature matching between photos for aligning and 3D reconstruction.
Affinity Photo JPGs performed "best", followed by Capture One, followed by the DNG, followed by sometimes darktable or openimageio depending on settings. In the few JPEG to TIFF output comparisons I did with identical processing settings, the difference wasn't significant. There may be a slight speed advantage when processing TIFFs (after rock conversion, not during), arguably offset by the storage penalty.
I could not identify exactly what steps made the best conversion, but I think it was a combination of demosaicing, histogram stretching/curves, and tone mapping/colorspace conversion. Based on relative performance differences when I tried to turn everything off in each piece of software (when possible) some fundamental (low-level) differences in raw post-processing approaches by the different software sppear to be the most significant.
Probably some of my limitations with the latter two were due to my ignorance and inexperience, but I tried hard particularly with dark table, to prove that assumption wrong.
Affinity photo produced point clouds with about 5-10% more matches and identical to slightly better accuracy than the DNG. Capture One ranged from slightly worse to at most 5% more matches than the DNG, with basically identical accuracy to the DNG. Using darktable with auto adjust only settings, or fixed settings (linear or gamma, base curve or no base curve), I was not able to get performance better than the DNG in both the 120 and 1000 image sets. The only time I could match or exceed DNG in the 120 image set was with settings that weren't truly fixed,And I don't think there was enough variability in the photos to truly represent "good"dark table did because this fell apart when I applied those methods to the larger photo set.
In contrast, affinity one beta in batch mode using default settings still beat the DNG by about 5% and was consistent in both photo sets. Capture one with auto levels adjustment only and the standard Sony A7R base curve also beat the DNG by about 5%. Both of these require the GUI, but I can automate them with auto hotkey.
JPEG s were saved at 100% quality, with floating point discrete cosine transform and no chroma subsampling where software permitted. The few tiffs I generated were saved in 8-bit LZW compression.
All of my post-processing tried to work primarily with raw processors, colorspaces, levels, exposure, tones, and base curves and used identical photo matching settings after the raw processing step. I didn't do any denoising, sharpening, lens correction, chromatic aberration correction, etc. Basically I was trying to minimize how much I did to the raw pixel data before processing with specialized software, and maximize the probability of getting photo to photo tie point matches, much like hugin/Panorama Tools, but for a specialized photogrammetry application made by Agisoft called Metashape (similar to PIX4D but with more knobs and dials), using an image processing technique called structure-from-motion, derived from computer vision work. Primarily we use these images to measure topographic change on rivers and coasts, but others do similar things with CGI, models for 3D printing, historic preservation, etc
Probably 70% of the conversions were using darktable, 10% Capture One, 10% Affinity Photo, and a handful using openimageio, Raw Therapee, and Sony imaging edge (?). I did some conversion runs with other software like DCRaw, graphicsmagick, imagemagick, gimp, Irfanview, and XNConvert, but could not easily figure out settings that would maintain white balance and auto adjust levels or exposure sufficiently,or produce images with high enough quality.
There was inconsistent honoring of the sensor crop between softwares, which was irritating, And technically frustrating because it means that the principal point of the image changes depending on what software you use. Which means that for structure for motion photogrammetry the same lens model cannot be tried with different conversions. For this reason I tried to do many conversions on the DNG because most software honored those dimensions more than the raw, but that wasn't true in all cases.
I was trying to dial in settings with all software that would optimize feature matching between images, and to develop a workflow that would make automatic adjustments to each image that was run through darktable-cli. Many of the runs used a single XMP sidecar file, before I realized that the adjustments were being made to all the images based on the histogram of the single image that was used to generate the XMP file (auto RGB levels, filmic, etc). Others just did auto levels, auto exposure, and a variety of base curves and color spaces.
If this is not enough information please let me know what I can expand upon. I don't know if I can attach images in this listserve so I won't try. There is a chart available at the Agiisoft forum post that I linked to in my original post, but maybe one that's a little easier to see on a post I made on the Affinity forum too:
____________________________________________________________________________
darktable user mailing list
to unsubscribe send a mail to [email protected]
____________________________________________________________________________ darktable user mailing list to unsubscribe send a mail to
