1.8 closed alpha feedback

[deanST]

Hi Mike,

yes, both are me: DeanST = DeanCN (not very original but easy to remember where I am!).
Thanks for sharing your process, no need to elaborate the log details. If you click the button with the </>
you get what I call "code blockers", which I assume are just inline opening and closing tags of a sort (html ? xml ?)
just put your log content sans masks between the inner two brackets.

Code: Select all

as an example ...

[Stefan B]

Hi Ivo,

just had a quick look at the new alpha. Worked like a charm and I really tried to cause a crash in SVDecon. Did not happen.

Sampling and calculating a PSF model based on the sampled stars yields good results for me.

Before:

crescentbeforesvdecon.png (371.83 KiB) Viewed 6758 times

After:

crescentaftersvdecon.png (393.05 KiB) Viewed 6758 times

Can one protect brighter stars from being deconvolved? Clicking the Mask or the ApodMask button leads to the same mask which is the ApodMask I assume. Is it possible that the general Mask contains the regions which are to be deconvolved, i.e. usually the whole image? Then one could use an inverted FatStars mask to protect brighter stars from being deconvolved.

Best regards
Stefan

[Burly]

Used 1.8 gpu version more times than not gpu version, not had any crashes win10 gpu nvidia 1060 card if memory serves me right .

[Mike in Rancho]

I too was up late and managed to then stay up even later after the new release came down.

Ran my M5 through it. Getting better (unsure if due to more familiarity or the new changes), but still not up to my 1.7 version of the image just yet. Though I did put a lot of elbow grease into that rendition.

It is appearing to me that all deconv is now global? I originally thought there might be two masks going on - application, and apod, based on the two buttons. But turns out the apod button really means auto-apod. The mask button allows modification of that apod mask. This worked well, as I was able to get a green-cored star within the glob by erasing other stars that fell in the blue box.

I did not try synthetic mode last night, so unsure how or if mask works in that. First release to me there seemed to be no effect, despite the language below the image as to green/not green pixels.

Again the global operation does lead to some difficulty, such that I am (so far) unable to strike the right balance of settings between the glob stars of my M5 and the bright star 5 Serpentis. It also still seems I am mostly getting brighter star cores rather than a coalescing or de-blurring of stars, but I will work on that some more.

That said, by adjusting spatial variance down and dynamic range up, I was really getting to some points where closely spaced stars were beginning to define quite nicely. Almost like seeing the Trapezium suddenly resolve out of a squarish blob. Totally cool, except in increasing dynamic range I lost the center of 5 Serpentis to a "gray hole." In the end I ran a layer blur on most star cores and 5 Serpentis to quash down any remaining artifacts and try (with only mild success) to handle the gray hole.

I'm also curious about the new correlation filter in Wipe. Is that parameter something objective we can shoot for? For example if I know my pixel scale is 0.9", or the seeing was probably at a certain level, should we choose to filter out detail below those amounts?

Mike

[hixx]

Hi,
There is a way to drive COMPOSE module into crashing....
Just load your LUM file and while loading wheel is still turning, hit the RED button.
GUI comes up with the directory, try load quickly. Now 2 loading tasks seem to be in the queue which brings whole Star Tools to a grinding halt...
No way to get out except quit. (process works with different colours too)
Probably ST should either enable parallel loading jobs or prevent users from starting loading jobs in parallel or just put them in a queue and execute these jobs one after the other, independing from "nasty" user behaviour
This bug was present in 503 and has not been resolved yet, not sure if this is already known

reagards,
jochen

[almcl]

Had a few goes with 504 and it seems very stable. Nice!

Just concentrating on the SV decon module and selecting stars in some M13 data I have from last month, and this is very minor point, but when trying to select 'green centred' stars, sometimes the blue boxes disappear from round previously selected stars and it is necessary to wait until both iterations of the new spinner (which I must say I like) to complete before the blue boxes reappear and one can see which items have already been selected and continue the selection process. But it does not always do this, and I am not sure if it's me going too quickly, or something else?

Has anyone else experienced this?

[Mike in Rancho]

Yes, sometimes the blue boxes vanish, luckily on a temporary basis. I'm pretty sure at first I was reselecting the same star. Can't recall if that acted as a deselect? I've avoided doing that since then and just wait it out when the boxes go off to hide.

I did save an apod mask last night (or at least I think I did). Haven't reloaded yet to see if it's just the mask that comes back, or mask plus blue boxes. Will try that tonight.

As for compose - hasn't that always frozen up if you don't wait for a channel to fully finish before loading the next, even in 1.7?

[bigjy989]

Notes on a quick run through SVDecon: No issues on Ryzen 5 & Windows 10
Dataset was 20x300s Ha 5nm. Stacked in Siril (normalization: additive with scaling) and 1-D polynomial background gradient removal prior to stacking --> bin 35% --> Autodev --> SVdecon (image pixel scale 1.1" --> 3.14" with 35% binning)

I found the best balance (+detail minus ringing minus spurious hot pixels next to stars) with:
Apodmask default. Extra Sensitive = too many spurious pixels after SV Decon.
sample area reduced to 9x9, point of which no additional difference in pre/post tweak.
#1: Spatial error needed to be reduced to 0.7 or so to eliminate hot-pixels around stars.
linearity cutoff fine between 50% and 90% - not so much effect shown in pre/post tweak.

#2: With the above settings deringing between 50-70% (100 is to much <35% no enough)
#3: De-ringing fuzz about 4x sample area seems best). 2x or less not enough.

With the above a lot of detail boarding on too sharp but the initial brightness of stars was maintained. A little ringing still remains around the stars (from extra sensitive mask, generally better balanced with default mask) but more deringing starts to blur the detail and become unnatural.

==> An option to blend the result with the original would be a good addition.

Before - Bin35 + Autodevelop:

After SV Decon:

Defaults SV Decon

[hixx]

Further Testing results..
SVDECON seems to run pretty stable in all modules except the Compose crash. Yes that might be historic, shouldn't happen anyway, surely there are ways to prevent this.
Except for the "grey hole" issue, SVDECON should be ready to go. I also had some runs on NBA and didn't hit any issue, but will continue to fry it.

"Grey Holes":
these seem to be slightly reduced in 504 but still visible. Actually these are precisely the red marked core regions in apodization mask. To mitigate this, can't the red pixels be used by the algorithm as singularity mask to not apply decon at all to those pixels, maybe using an additional 1 pixel fuzz?

Here's my "wishlist":
- PRESETS for 1.8 (fast/mild and precise/strong), 1.7 and (probably) 1.6 so users will be provided a good baseline for decent results of the relevant algorithms. If a user would like the "old" or "better" DECON tool, it'll be at his fingertips too.
- APODIZATION MASK 1: Show only green (maybe yellow) core stars. This provides better overview and prevents "bad" sampling
- APODIZATION MASK 2: Fade out tiny stars below 6x6 pixel circles (these just fit into smallest area selection). The screen is swamped with fiddly but unusable stars marked which do not add too much benefit. Again more overview and less prone to user error.
- APODIZATION MASK 3: display number of selected stars
- AUTO SAMPLE SELECT feature: Automatically select all green core stars of at least 6x6 pixel sitting before dark background and without extreme deformation (double stars etc). This would provide a great 1st run result the user could build on top. It's probably a frequently asked one as well.
- SAMPLE AREA1: is it possible to use a circular sampling shape rather than a square to reduce neighbor star sampling effects. Havent come across a square star yet anyway
- "SAMPLED AREA" control to auto-adjust to largest selected star. This would prevent user error and poor results by either selecting too much or too few area. There is no point in having a 15x15 default if the largest selected star was 9x9 and neighbor artifacts would impede the result
- INCLUDE ONLY control on SVDECON. A tool similar to the one on the mask, but driving DECON algorithm... This might prevent issues from selecting stars in nebulosity regions. Images of large scale nebulosity might barely provide any dark background regions eligible for sampling. If stars on top of (not too bright) nebulosity could be selected but DECON algorithm could still handle that would be great. Having said that I did not see any weird effects selecting stars on top of nebulosity yet as long as it was not too bright.

regards,
jochen

[bigjy989]

Getting closer with tweak but still struggling to eliminate all artifacts in the SVDECON module: Links to are the private AB staging area.
Regular decon still working well, SVdecon is great resolution but hard to fully eliminate halo's and grey cores.
Full resolution process (a little too much for the data but confirm no crash) (~24MP) with the same 20x300s Ha subs from the previous post.

Revision E (through denoise) & F (SYNTH to attempt to correct halo & grey cores) added

E:
Description: Revision E - full process at full resolution. Autodev - 33% on linearity, 3.3 pixel dark anomaly, 0.9 sensor detection. HDR - 2000 pixel strength 1.1 Sharp - default mask, 500% strength, 33% dark/light bias; SVDecon - 90% derining, 40 pixels, Spatial error - 0.67, Dyn Range Extension 1.05, Linearity Cutoff 80%, Sample 15x15, Iterations 20 -------------------- Life 90% DimSmall -------------- Denoise 3.5" --> background noise ok, still some ringing and grey holes in stars Edit

F:
Description: -------------- Revision F - attempt to fix stars (grey holes and rings via SYNTH) SYNTH - refractor sample size 4000 same mask as sharpen minus nebula highlights (via lasso), Core whiteness 100%, Blur 2 pixels, blend 0%, brightness 20%, core noisiness 10%, Image Dia 220 arc-min

F is getting really close but still a dozen or so noticeable artifacts that need to be corrected with heal / clone stamp.

I'm still feeling that a blending option is a necessity. The added detail is great but a light hand to the limit of artifacts is a must.