2019 September: Astro Pixel Processor and iTelescope.net celebrate a new Partnership!
Challenging milky way panorama - call for assistance (1.074.1)
I'v made a 34 field 620 exposure mosaic of the Milky way with a Canon 6D. Since 620 exposures are too much to register in one go for my computer I first did a test run with a single exposure per field and got a nice result out of that. For memory and speed reasons I reduced the result to 0.2 times what it could have been in.
So far so good.
Then I've combined all exposures (typically 19) of each field to a single intermediate stack with the intention of doing the same as with the single exposures. And that's where I've run into serious problems with the registration step. Anything from error messages to very creative and artistic stack results with extreme distortion. I've played around with the #stars in the analyse stars step and with just about all the settings in the register step, to no avail. Since each attempt took many hours of processing I quickly decided to combine multiple fields into larger intermediate stacks, reducing the number of files to be combined to 7. Also here, for memory and speed reasons I chose to reduce the scale of the stacks to 0.5 of the original size.
And that's where I'm stuck now. I can't get all frames to register properly. Some of them will register, but of course I want all frames to be included in the final stack.
Along the way I've seen various error messages pop up, mostly this one or a similar version thereof:
"APP couldn't register any frame to the reference frame
Please choose another reference frame or,
adjust star analysis paramters to increase the
amount of detected stars. "
I'd like to tap into the collective wisdom of the APP users to help me get my final stack.
The lens is a Samyang 135mm, a very sharp lens that in the original images already gives close to undersampled star images on the DSLR detector. The reduction in size by a factor 2x2 of the intermediate stacks makes the stars even sharper so that may be one of the reasons for the difficulties I encounter.
Has anyone else been where I am now and could help me with the last steps? Any words of wisdom? Smart solutions? Crazy ideas that will still run in 16 GB of memory and reasonable processing time?
In the mean time I'll struggle on. Grateful for any help here!
Next attempt: registration with the single exposure stack from the picture above, so the remaining stacks have a guideline to help them along. The result:
Encountered error in module:
This error I've seen multiple times. The settings were with 10000 stars per frame, pentagons, and dynamic distortion correction. What I think is happening is that due to undersampling some of the originally bright stars become faint, and due to the large amount of stars in the milky way the algorithm finds matches anyway, but not in the right spots. This leads to immense distortion (if it still fits in memory and other limits) or error messages because an algorithm is presented with input that's totally crazy.
Which settings should be robust against messed-up star brightnesses and (totally) wrong identification of star patterns?
With 500 stars I get the following familiar error:
Registration failed on some of the frames !
Possible solutions for a successfull registration are:
- increase the automatic #stars target in 3) ANALYSE STARS
- disable same camera and optics (if enabled)
- try different pattern recognition descriptors
- increase the scale stop level
especially if you have data of different image scales or
with little overlap between the frames
- enable flip descriptors in X/Y
if you need to register data from different telescopes
Slightly different settings, 10000 stars, any pattern recognition descriptor, DDC, normal instead of mosaic registration, and manually selected the single exposure stack frame of the milky way as reference:
Encountered error in module:
So it may be a good idea to try the latest beta as some of the java errors were tackled in that. Having said that, these large projections are not very well supported yet, Mabula is working on more for these kind of tasks, not sure about any timeframe for those @mabula-admin
Hi Ralph @ralph,
Yes, you have a big project here... lots of memory is needed.
In my experience, 500 stars is too little, but 10000 stars is not slight different, rather it is way too much in any case that I have seen.
Try 2500-4000 stars with quads and lower distortion margin. I have reduced the default margin in APP 1.075 since it is fine even for data like this.
If quads don't work, try triangles.
Oh and please check this:
Some remarks, tips:
- The most efficient and most robust way to register any data set as a mosaic is to first integrate the individual mosaic panels. Then register the mosaic using only the individual integrated panels.
- This holds for other mosaics as well. Don’t try to create a mosaic out of 100s of indivdual frames. First create the mosaic panels.
- To prevent integration/stack artefacts at the borders of the mosaic panels, use both Multi-Band Blending (MBB) & Local Normalization Correction (LNC) to reduce the amount of integration artefacts considerably.
- The actual Mosaic integration needs to be done, using MBB & LNC as well to correct illumination differences and to prevent the well-known mosaic seams.
- Regarding: Dynamic Optical Distortion Correction.In most cases, you can leave this disabled for the individual mosaic panel integrations. Only enable it if registration RMS is higher than 0.5 pixels and registration is visibly not correct. In the mosaic registration, you really need to have this enabled always.
So for your final mosaic, first create the individual panels, that is much easier, faster and much more robust 😉
In addition, use calibrated projective model and use the equirectangular/mercator projection with adjusted x/y move of Center Of Projection. You can check on the reference frame what this does with the l-c-registered image viewer mode 😉
Main developer of Astro Pixel Processor and owner of Aries Productions
OK, all problems solved with the latest version, 1.075! Thanks for the update, this is a big step forward!!
What I did, using the lessons learnt in my earlier attempts, is the following:
- Rotated and cropped away (most) of the horizon in those exposures affected
- Quite minimal background correction for the exposures close to the horizon
- Stack the individual mosaic panels (between 9 and 20 exposures per panel) with Super Pixel (to reduce the size of the final mosaic and speeds things up along the way); 500 stars (default setting); for registration same camera, no DDC, scale start/stop at 1/5 using quadrilaterals; regular mode for normalization; 1st degree LNC with 3 iterations, MBB at 5%.
- Load all 34 mosaic panels
- 2300 stars
- Quadrilaterals, scale start/stop at 1/10, DDC, no distortion model, registration mode mosaic, calibrated projective, eventually manual selection of the reference frame, Mercator projection, 30 degree rotation (which depends on the reference frame chosen), and so far no change in the CoP position
- Advanced normalization
- Average frames for the integration, 2nd degree LNC with 2 iterations (needed for the somewhat crude background correction done earlier), 10% MBB, no rejection
- I did a couple of test integrations to see how the settings would work out. These were with scale 0.1 (nice and fast and minimal memory requirements) and no LNC (also faster). I haven't quite figured out how to tune the Mercator projection settings. Rotation in the 4) Register tab is not fully consistent with that in 9) Tools -> batch rotate/resize, after a couple iterations I got the result I wanted.
- Just to be on the safe side after all this work: before final Integration I saved the Normalized results, 2 GB per file, 29712 x 5637 pixels. That's big...
I'm about to press the "Integrate" button now with the settings for the hopefully final integration. Wish me luck 😋