Twenty four hours pas new year’s eve, I was out for my first photography outing of 2019. I wanted to take some starry sky photos, and decided to go to the Pointe des Châteaux area. My goal was to take pictures of the Milky Way while having some interesting foreground.
I used the photopills application on my smartphone to plan my trip and see where the Milky Way would intersect with the horizon so I could get both the Milky Way and some point of interest in the same frame. The POI I wanted to feature is a big concrete christian cross.
I entitled this post “The Creations” because in the same picture I see the duality between the religious (here Christian) and the scientific viewpoints on the Creation:
just above the foreground hill, at the bottom of the Milky Way you can see a reddish area: it is called the Carina nebula, which is an H II region, in which stars are forming along with protoplanetary disks;
if you let your eye follow the hill to the left, you will see the 10 meters-high concrete cross on the clifftop, which was first erected by catholic Christians in 1946, and replaced by the one you see here in 2002.
The technical stuff
This image is a stack of many different frames, first processed in Siril 0.9.10-rc2, using:
Siril demosaiced and processed the raw files (Fujifilm X-T2) directly, which is great (support for x-trans raw files was added earlier in 2018). I just used one the scripts posted on their website, which takes care of everything.
I ran the script twice : once with star registration, to get the stars aligned, and then without registration for using the stacked foreground.
The 2 resulting 16-bit TIFF files were slightly processed in RawTherapee (exposure, tone curve, contrast, white balance).
The final step was to merge both TIFF’s in the GIMP (masking the foreground of the star-registration image, manual removal of hot/stuck pixels, ‘creative’ midtones dodging in the foreground, cropping and resizing to 1920x1080).
that is some fantastic work. I was looking for ways to improve my night/astronomy shots and it really looks like I should give SIril a try! thank you for that…
The more I look at this, the more questions I have. What are “bias (offset) frames”? How about light vs dark? They seem to be at the same exposure (iso 800 10"), why makes the dark frames darker? f/ stops? What about “flat field frames”?
I guess you used a tripod for this, right? Did you also use a remote? How long did you wait for the camera to settle? Does it matter? I calculate you took at least 60 frames at around 10 seconds each, which means at least 10 minutes delay, during which time the whole sky will have rotated quite significantly (2.5° to be precise) - is that what Siril deals with? How does it deal with the “foreground” (the earth landscape) which doesn’t rotate?
Thank you so much and sorry for flooding you with questions…
Thanks anarcat.
Regarding the different types of “calibration” frames (darks, flats, bias), your can have a read here. Regarding the number of frames of each type needed, there’s still some debate. Because people asked me on dpreview why I took so many calibration frames, and that I just used what I saw on websites, I thought I could do some measurements myself. Below is a chart of noise level measurements (done in Siril) as function of the number of dark or bias frames stacked:
Based on this, I think 20 dark and flat frames are a good number, 10-20 for light frames, and 100 for bias frames. Bias frames don’t take long to be made. Also, when Siril has run the whole process, “master” calibration frames are saved, so you can delete the original frames and just cut the single master frames which contain the information you need, and you can reprocess your light frames at will using those (even for other shootings, as long as the shooting conditions don’t change significantly).
I took only 10 light frames at 10 seconds, using either the integrated intervallometer or the 2 sec. delay timer, on a tripod, with just 1-2 sec of settling time. My camera is mirrorless, so there’s no risk of mirror-induced vibration.
Indeed there’s a visible rotation of the earth, so I ran Siril twice: once with registration (i.e. star detection and aligning), and the second time without alignment to have a stacked foreground. Then I merged the two resulting images in Gimp, using a hand drawn mask, to have both stars aligned and a sharp backround.
Since my original post, I’ve learned more stuff with Siril, and reprocessed my frames. Here’s the current result, this time with different tweaks in Rawtherapee, and without dodging in Gimp, just masking the foreground:
Wow, okay so that’s more involved than I expected. The FAQ’s TL;DR:
Light frames: normal pictures with the stars in it. those get stacked
Dark frames: same settings, but with the lens cap on (10-20 shots)
Bias frames: same as dark frames, but at fastest speed, e.g. 1/8000s (10-20 shots)
Flat frames: same focus as light frames, but overexposed (10-20 shots)
Those numbers are according to the FAQ, but you seem to show there are still good improvements to be made above those numbers, particularly for bias frames.
Thanks for the hand-holding! I guess I got more reading to do as well…
Oh and congrats on the new rendering, it does look much better, especially in the upper-left corner, where the smudge is totally gone…
Regarding the mushy corners in the first version of my image, I think the trouble comes from the heavy vignetting introduced by my lens at f/2.0, which requires a heavy correction by the flat field frames, making noise more visible if I try to render the scene brighter.
Based on my experience with non-photographic (engineering) data: Noise amplitudes from addition of unrelated sources behave as the root of the sum of squares (\sigma_{tot} = \sqrt{\sigma_a²+\sigma_b² + \sigma_c² ...})
This means that e.g. the bias frames, although you can get the noise on them a lot lower with more samples, wouldn’t actually help that much to reduce overall noise, and that adding more dark and light frames might have a much stronger effect on the end result even if the relative improvement of dark frame noise is not spectacular.
That said: That’s for unrelated noise sources, and for simply adding signals. With the exception of dead pixels or such, the noise in these frames should be unrelated (local noise amplitudes may be related, but not whether a given pixel in a given exposure will be higher or lower than that pixel’s mean expected value). But of course, Siril does not simply add the data from dark, bias and light frames. There’re some multiplications, divisions and subtractions going on as well, so some of the noise sources will be amplified and some reduced.
I guess the best way to tell how much the sample numbers affect the end result is therefore either to take the curves above and follow the maths (if you know how to do error propagation and also what Siril does under the hood), or to have some uniform surfaces in the light frames, (e.g. slightly defocussed grey cards, some kind of test chart or something), and then check out the noise levels in those parts, while varying the number of samples used – that’d mean you’d have to shoot your light frames twice unless your scene already has naturally-occurring uniform areas. At least what you’d learn would be valid for all similar scenarios.
To my understanding, the flat frames should not be overexposed – they should be auto-exposed (better: ETTR but not blown out) photos of an evenly-illuminated surface. The idea is that if your camera were perfect, you should get a picture of one single solid colour. They’re used to correct for vignetting etc…
As far as I understand, they should also not be particularly sensitive to noise since they get a lot of light and thus have a fairly good signal-to-noise ratio (SNR). Since the light field is divided by the flat field, that signal-to-noise ratio (after subtraction of dark field noise) is more or less directly applied to the light field. If your scene is a night sky, then the SNR in the light frames will be much worse than that, and should receive a lot more attention.
When I did analysis of microscopy images a long time ago (without dark/bias frames), I’d simply blur the flat field images, turning them off as a source of noise altogether. That’s fine because there should be no small-scale information in them. Of course, Siril may already be doing that internally, and it may be applying the info from Dark and Bias frames to them, too, so supplying pre-blurred flat fields to Siril may or may not be helpful.
What are “bias (offset) frames”? How about light vs dark? They seem to be at the same exposure (iso 800 10"), why makes the dark frames darker? f/ stops? What about “flat field frames”?
