StarAlignment has trouble finding a transformation unless I stretch first

[Nocturnal]

Hi,

I downloaded a DSS image from ESO to be template for placing DBE samples. The image shows plenty of stars but staralignment can only find a transformation if I first do a linear HT where I discard the empty black and white ranges of the image.

Is this to be expected? I can upload images if needed but they're rather big so I'd like to hear the theory first.

[Juan Conejero]

Hi Sander,

No, of course that shouldn't happen. In fact, it is very weird. And yes, I need to see the image... But you're lucky because I can download the DSS frames if you tell me their coordinates

[Niall Saunders]

Hi all,

I was wondering whether an extract from the GSC catalogue might not be a better way of creating a 'reference' FITS image? At least it would be clear of background noise.

[Nocturnal]

Hi Juan,

I'll look up the coordinates I was working with and let you know.

Niall: it's not about the stars but about the nebulosity and I don't think that's present in the GSC image you're talking about, right?

[Niall Saunders]

Hi Sander,

I realise that - and that was why I wondered if a 'false image' made up from star locations and magnitudes (extracted from a suitable star catalogue) could be used as the 'reference' frame in the Star Alignment process - in much the same way as you would do 'astrometry' in the likes of AIP4WIN, etc.

Then, once all your 'real' images were aligned to the 'false reference', you would just dump the reference frame prior to stacking.

In my mind, and with a very simplistic analysis, the 'perfect' stars from the 'star catalogue' at least give PI a good 'head start' for trying to locate the same stars in the 'real' images. But, perhaps I haven't fully thought this through - maybe I am missing something 'obvious' here. It wouldn't be the first time 

Cheers,

[Cheyenne]

Just a quick comment...  -- you might want to try the FFT registration script that's out there.  If there is enough nebula structure the FFT registration should be able to sync up the images, as opposed to having staralignment looking for specific points to lock onto.

[Nocturnal]

Hi Sander,

No, of course that shouldn't happen. In fact, it is very weird. And yes, I need to see the image... But you're lucky because I can download the DSS frames if you tell me their coordinates

I can't reproduce it now with the frames that I have. If it happens again I'll upload the images right away. It happened several times so I don't think it was a fluke.

[Nocturnal]

Well that didn't take long

Please find a 16b TIFF image I use as the reference and a gz of an ESO DSS image here:

http://cid-a93625fef5ca95fb.skydrive.live.com/browse.aspx/PixInsight/StarRegistration?uc=2

Using default settings I can not register the DSS image. Once I discard unused white and black space registration works fine. In fact the stock image apparently renders only 5 stars even though I see plenty. Removing the band at the bottom it only finds 2 stars.

[Juan Conejero]

Hi Sander,

Thanks to these images I have been able to fix this bug very quickly. The next version registers these images without problems: with default parameters, StarAlignment finds 163 star pair matches in two attempts (the first attempt always fails because there is a strong difference in scale between the images). Besides fixing this bug, I've also improved the star detector. It is now more robust and finds less poor-quality stars, which in turn improves the performance of the star matching algorithm.

Thank you!

[Juan Conejero]

An interesting follow up. The DSS image that you have uploaded (which is obviously a mosaic of two or more DSS plates) has a strong distortion that cannot be modeled as a homography. For some reason, the left half of this DSS image is "compressed" toward the center some four pixels on average, with respect to your image. This explains why StarAlignment is unable to find any star pair match on that half: there is no way to fit such large-scale distortion using a homographic projection. What is wonderful is that the star matching engine automatically finds the best possible fit by choosing star pairs only on the right half of the image. RANSAC and linear algebra do these things

[Nocturnal]

Great! Glad to be of help.

[Nocturnal]

This problem is still easy to reproduce in 1.5.5. I can upload images again.

[Juan Conejero]

Hi Sander,

This is strange. I can register your two images with 1.5.5 and default StarAlignment parameters:

http://forum-images.pixinsight.com/legacy/1.5.5-preview/StarAlignment01.jpg

Are you working with the same images? Are you sure there is a valid (translation+rotation+scaling+small-scale-distortion) transformation between your two images?

[Nocturnal]

Hi Juan,

no, these are two new images. Yes, I'm sure a transformation exists a) because I can see the two images match and b) because after I move the black and white points on one of the images the transformation succeeds.

I'll upload the two images later tonight. You can see the result (Pelican) in the gallery.

[Nocturnal]

Try registering the black and white image to the color one

http://cid-a93625fef5ca95fb.skydrive.live.com/browse.aspx/PixInsight/StarRegistration2?uc=1

Now use the histogram process to remove empty white space and bring the black point up. It will now register.

I suspect the large black areas due to the mosaic operation I did to create the image throw off your statistics.