3 ms·
I think this is the paper: https://arxiv.org/abs/2105.01056 https://arxiv.org/abs/2105.01056 . It's all open sourced, and Python. From the paper, previously yo
by lostdog 4y ago
I think this is the paper: https://arxiv.org/abs/2105.01056 https://arxiv.org/abs/2105.01056 . It's all open sourced, and Python.
From the paper, previously you needed a sequence of images from a single night to generate a "tracklet," which lets you estimate a guess of the orbit of the asteroid. The guess can be compared with images from other nights, to confirm the existence and orbit of the asteroids. Without the tracklet, the asteroid could be heading in any direction, meaning that the set of possible orbits is huge, and there would be many many possible matches to other images.
The key insight of THOR (their algorithm) is that even a mediocre guess of the orbit makes the true orbit stand out. From an observed object, they predict a test orbit. Then they take all other images that could be along that orbit, project all objects in each image to the original image according the orbit, and then stack the projected images together. If their guess of the orbit was perfect, there should be a clear dot in the stacked projections, where that single asteroid was seen in all the images. But even an imperfect guess is still useful. Any object that's somewhat close to the guessed orbit will show up as a line in the stacked image. So find the lines, use some math to refine the estimate of that objects orbit, and now you've got the object.
(I'm not in this field, so take all explanations with a grain of salt please).
- kadoban 4y agoReally interesting. They're using a Hough transform, which is used to pick out shapes from an image (lines and circles most often), especially popular before ML techniques for object recognition existed. Sounds like a great way to search for asteroids, never would have thought of it. It's essentially just a coordinate transform that lets you add up evidence for a particular object, and the objects stand out as little bright points. I'm sure they're doing lots of clever things on top of it of course though.
- dylan604 4y agoLots(most?) comet discovery is now done off of sequences of images and looking for the dots that move in relation to the background stars once the stars are aligned. At that point, you could do differences with subtraction between the plates so that aligned stars disappear and only interesting dots remain. No machine learning required.
- salty_biscuits 4y agoHough was looking for the tracks from charged particles in cloud chambers, so the algorithm comes home after a while I guess.
- kadoban 4y agoI did not know that, that's a nice origin story!