Lights in the Night Sky

Three minutes of arc and thirty minutes of time of the sky over Siding Spring, Australia, on the night of 5 March, 2016.

Long have I wandered in the dark spaces between the stars, and in the vaster spaces still of all the ways you can screw up thinking about them.

My problem has been that I've been trying to come up with a way to find moving objects without ever actually making a judgment that says This Is An Object.  I don't mind writing posts about efforts that don't pan out or ideas that need to be corrected later, but I'm kind of shy about writing when I know darn well that I am totally full of shit.  Well, I believe I have made some advances lately which have persuaded me that I am now down to about half full, so here we go again.

You can, of course, get better science writing elsewhere, folks -- but not at this price-point.

The two great principles that I learned from working in machine vision are:

• throw away as much data as you can  and
• put off binarizing decisions as long as possible

By throwing away data, I mean:   the three images I took tonight come to about 60 megabytes.  If I find paydirt in them, it will be a little moving object that can probably be described with a few small numbers.  Maybe 32 bytes total.  A good way to understand your algorithms is that they are a way of discarding the useless data, without getting rid of what you want.

I think it was Michelangelo who first said:

Questo scrittore è un idiota. Vorrei fossi ancora vivo, così potuto fargli il culo.

 which of course means:

I find asteroids by getting rid of all the pixels that they ain't.

By put off binarizing decisions, I mean: a decision like these pixels are foreground and everything else is background -- kind of pins things down.  Even if you preserve the actual pixel values, all your algorithms are now working with that binarizing assumption.  It's best if you can find a way to refine your data-ore as far as possible without making limiting judgments like that.

A rare photograph of the T27 telescope in its natural habitat.

Well, I've tried to stick to those ideas for a long time and they have led nowhere I want to be.  So at last I emerge from the wilderness with this insight:

You need to find the objects in each image.

There is just no way around this.  You need to make a binarizing decision that These Things are Objects, and the Rest of the Image is Just Background.  There is simply no way around this.  You find the objects you are interested in, do that in each image of your sequence, and then look for objects-that-move.

Seems pretty obvious, right?  Yeah, well.  It is possible to get Way Too Fancy while thinking about this stuff, and that is maybe the best way you could find to never get anything done.

There is one more insight that is at least as valuable, about how to do the vision work.

The night sky has many different kinds of objects in it.  Use a different algorithm for each.

So there you have it.  This gives me a way to proceed.  I don't think I will be wandering anymore.



Thanks to the heroes at iTelescope who let me rent T27, and to the heroes at PlaneWave Instruments for making it!