astrometry.net fails for images with small number of stars

[withboobs]

Dustin, thanks for the great work,

Now a question. I have noticed that the plates fail to get resolved if they have small number of starts on them over wide angles. Case in point is, e.g., submission 1306496, http://nova.astrometry.net/status/1306496, which contains 29 bright stars recorded on a plate  with 48 arcsec/pixel resolution.

Is there a hope that such plate can be resolved with nova.astrometry.net solver?
I did try some of the suggestions given to other users having similar problem (downsample=4, tweak=1) none of them works.

In addition I have tried rotating image in 15 deg increments, and increasing positional error to 20 and higher values, all without success.

It may also be possible that the plate contains either non-existent starts (can those be identified?), or that itself is a mashup of other plates (if one tries to resolve sub-plates what is the minimum number of stars on the plate?).

Regards,

[Dustin Lang]

Hi,

You might try submitting a sub-region -- a closer-to-square region.  Long thin images can be problematic.

Are you submitting text x,y coordinates?  In that case, downsample has no effect -- that only applies to images.

Rotating the image should have no effect -- the Astrometry.net algorithms are all rotation-invariant.

A few extra stars are no problem.

cheers,
--dstn

[withboobs]

Thanks for the quick reply.

The data set comprises of xy table extracted by source extractor from an image. From source extractor I also get the magnitude of stars, but that does not appear to be important for the web interface besides stars being ordered in the table by their magnitude (from brightest to not so bright).

The reason why I considered rotating it was because at 180 deg rotation the solver actually reported identifying some of the stars, albeit at 3 times smaller scale (image at 48 arcsec/pixel, reported 16 arcsec/pixel).

I was reading the report you put on arxiv.org about the method, and was wondering if quads are parity invariant (mirror image)?

[Dustin Lang]

Hi,

Excellent question -- the quads are not parity invariant, but by default we check both parities.

cheers,
--dustin

[withboobs]

Dustin,

from reading post on arxiv, it transpires that the solving should be irrelevant of the rotation and parity, Yet it is not.

Now, if I read it right, by adding more stars the odds should go up because the number of false positives decreases. This I haven't seen. E.g., when I increased number of stars in the images from 29 to 51 on average, the odds did not go up between jobs with one and the other number of stars.
Any advice on that topic?

Then, if I start guessing where in the sky the stars on the image are, would this improve odds to the point of software being able to identify the stars once I guess the right RA and DEC?

Regards,

[Dustin Lang]

Could you please send a link to the rotated example?  (The one where you got a wrong solution with 16"/pixel)

We report the *best* odds ratio found while testing stars.  Adding more stars is not guaranteed to increase the odds -- for example, if there are only 20 reference stars, than once all those have been matched, the odds ratio will not increase no matter how many more stars are added.

Knowing the RA,Dec does not increase the chances of finding a match, or increase the odds once a match is found; it only makes the search faster.

cheers,
--dustin

[withboobs]

Dustin,

thanks for the continuing interest in this topic.

What I am working on comprise between  21 and 51 stars in XY format provided by source extractor (c/python library sep), cf. submission 1304079 and 1319900, respectively.

I have a collection of between 21 and 51 objects in pixel- XY coordinates on the plate with nominal resolution of 37 arcsecpp.
Initially I choose left or right edge, and from it remove increasing number of stars while fixing the plate resolution between  35.9 and 38.11. The mirror image rotated by 180 deg achieved so far the highest log odds of 7.1e6 (submission 1310687, 4 matrched objects) of all my attempts altogether.

Then I changed the way I am submitting the tasks. I now set
(a) one of 4 orientations: as is, rotated by 180 deg, mirrored, mirrored and rotated by 180 deg.
(b) plate resolution arcsecpp in the range 16 to 40 in 0.5 increments

This on its own did not work with best odds below 1e6.

Then I introduced sliding window (parameters length of the window, and the offset). In the normal orientation I got three different groups of matches for the windowed star subgroup:
one at 19.5+/-0.5 with 7 matched stars, at 17.0+/-0.5 with 6 matched stars and 22.5+/-0.5 with 5 matched stars. From reading through the logs the group with 6 stars is around RA,Dec~30,30 while the group with 5 stars is at 80,-70.

Would varying positional error improve odds of finding the solution.

[withboobs]

Dustin,
would it be possible for you to resolve the plate for me? I can send you the csv data file, and you tell me all the tricks you had to do in order to resolve it.
Regards,

[Dustin Lang]

It sounds like you have tried a bunch of things.  I don't have additional magic to apply to the problem...

The things I would check are:

- are stars being detected in a reasonable way -- are all the brighter stars being detected, and are the reported centers correct?  (saturation and other issues can cause stars to be missed or mis-centered; getting these right is vitally important)

- are the pixel scale limits you are giving correct?  With digital cameras it's easy to miss a factor of the chip fill factor (eg, my Canon has a small chip, so a 50 mm lens effectively gives the field-of-view of an 80 mm lens.)  The scale limits are treated as hard limits in the code -- if you set the limits to be 1 to 2, then a quad that would imply a scale of 0.999 would not even be considered.

- I would crop to nearly-square (say aspect ratio up to 1.5) image subsets

- does this camera produce a lot of distortion?  If so, this often makes it difficult for us to get a solution; we assume a flat focal plane, so large distortions can cause failures.  Often the distortion is not as bad in the middle of the image, so maybe try cropping out a central region.

- what filter/bandpass are you using?  The 2MASS index files we provide for download use J-band to order the brightness, and the nova.astrometry.net site uses USNO-B Red bands.

cheers,
--dustin

[withboobs]

The plate resolution is given, and is based on the camera focal length and its film frame size. Yes it was an old SLR.
Going throught the numbers gives pixel scale 37 +/-1 arcsec per px. At that scale no matter what I do to the sources I get no solution, with best odds below 100K.
In submission 1347573 all seemed well, and 6 out of 7 brightest sources were said to have been identified, then solver like it forgot about it and ended not solving the plate.

From what you are saying, what can I conclude if I cannot solve the plate? Can I conclude that instead of real stars it shows an artistic rendition of stary sky?
And if I claim that, can you prove me wrong by going over the data set and solving it? In other words, nobody will trust me if I cannot solve it, but they will trust you.

That was just a thought. I hope I am not crossing my boundaries.

Regards,

[Dustin Lang]

Can you send an image?  It's hard to diagnose issues with only source lists...

cheers,
--dstn