How a Programmer Can Discover an Asteroid | Hacker News Reader

How a Programmer Can Discover an Asteroid

ianww.com

65 pointstyppo11 comments

Interesting. Here's the link to the Asterank Discover page: http://asterank.com/discover

I've gone through a handful or two of images now. Although he says "The app occasionally serves control images to get a sense of whose responses are trustworthy," I have yet to see anything move. It'd be nice to have a higher number of catch trials with a game-like interface. "Congrats, you found the moving dot! This is asteroid xxxx" or: "Oops, you missed it, try again" or: "Hey, this is a new one! We'll check on it and get back to you."

I think that'd help train and reward those who look through all these images.

Edit: after a few dozen, I think I found one. But I don't know if it was a catch trial or not.

Agreed. The lack of feedback is slightly troubling when you think you see one, as is the inability to reposition or remove a dot. Still, it's quite a bit of fun.

This seems like a good time to point out the Zooniverse project (https://www.zooniverse.org/) which is using similar crowdsourcing techniques in the browser to do all kinds of things from cataloging galaxies to entering data from ships logs to extend our view of earth's climate.

The 3d view, http://www.asterank.com/3d is very snazzy.

Quite difficult (but very cool).

The way that the brightness (if that's the right term) varies between frames made it hard for me to distinguish between movement and adjacent objects appearing and disappearing. I wonder if there would be a way to normalize the brightness based on the brightest of the frames (i.e. the frame with the most objects visible) to make it easier to detect actual movement as opposed to what looks like movement due to stop-motion-like hiding and revealing of adjacent objects.

Hey great project I always wanted to set set aside time for similar projects, mainly going over through the troves of data available from the Kepler that is publicly available. I might start a a project and put it on github but as you all know such things are incredibly time consuming, yet are possible with a small and dedicated team.

If anyone would like to collaborate on such a project in a dedicated and professional manner I would love to chat.

I have 2 friends onboard (google and MS search engineers) who will put in time too given proper dedication.

I wonder if you feed a few of screenshots of the video into a Reduced Boltzman Machine if you could get it to come up with what it thinks an astroid looks like. seems like it would be perfect for this kind of thing.

mmm a "zoo" project yeah, crowdsourcing science like this is actually pretty popular. They use it for determining galaxy types and finding planets... unfortunately it's often highly biased once you get to the realm of Machine Learning failure, you end up in a regime of false signals that can easily fudge amateurs

like galaxies, crowdsourcing science found that that most (like a significant portion of most) galaxies spin clockwise or something, but it should be purely random. This was a big blow to the galaxy-zoo project from like 8 years ago. crowdsourcing science is hugely powerful, but I don't think it should be tackled or implemented by amateurs really... crowdsourcing should really be in the form of spare CPU cycles like BOINC projects...

Although. Asteroid detection does basically operate on the algorithm this guy proposes. emphasis on _basically_ there are literally thousands of types of variable objects that can flicker out and thousands which can flicker in. and a pure random background which gives strong positive signal on an instantaneous timescale. what he's doing is basically 1960s tech with a million grad students.

I also want to blow a load on this guy's burger for this fucking statement: "A year ago, I set aside my doubts and started to innovate in the space industry as a complete outsider." Give this fuck 3 hours at CfA and see how confident he is of innovating anything. The real people (CSCI People) changing astrophysics are the deep machine learning guys, the data structure guys and the mapping algorithm guys. not the bloggers.

This is a worthy idea and a lovely example of programming (particularly web development) tackling interesting problems outside of your domain...I think there's huge potential for programmers who, arrogantly or not, attempt to tackle industries and fields as outsiders.

However, I don't understand why the OP believes asteroids too small to be seen through computer-vision algorithms would be noticed by human eyes? I'd think computer processing would be much better (and configurable) at detecting the smallest cases and, at the same time, be better (with some additional tweaking) at reducing false positives. I'd be interested in seeing whatever data/edge cases the OP found that led him to go this crowd-sourcing route.

(OTOH, bringing people in to help find asteroids is a great way to bring attention to the project...but hopefully he's doing a lot of pre-filtering/processing of the images to serve up just the most likely candidate sequences for humans to go over...and if so, what's the threshhold he's set for the pre-processing)

Computer algos do a good job at asteroid id'ing - they discover a few every day - but most experts I've spoken with (including people running large sky surveys like PanSTARRS and CSS) agree that crowdsourcing is an approach worth trying. I've updated my post with a few of the reasons why, which generally relate to the unreliable quality of sky survey images.

From personal experience, the algorithms and data used by sky surveys are not readily available to the public. This inhibits innovation. Without a PhD, there is little opportunity for me to build on state of the art and improve them directly. I agree that with the images I've collected, there may be some interesting algorithm/ML opportunities. Part of the reason why I started with crowdsourcing is to get a decent training set.

I appreciate the feedback!

Thanks for the reply and update! Also, I just realized you have a link to the source repo, which would answer a few of my questions. If there are certain conditions in which false positives arise frequently from image analysis, perhaps a refinement of the parameters (if photo_taken_during_hazy_night && by satellite_in_position(x,y)_relative_to_viewport, then increase fuzziness_factor) might also cut the noise. From my layman's perspective, the problem seems to be "easy" because of the relative simpleness of the shapes (whitish dots on a dark area) and expected movement patterns...And so the fact that it isn't means that there's other problems (data quality) or a major flaw in (my) assumptions...which is even more interesting.

Conversely, it'd be interesting to set up a measurement-framework to test which conditions are users most likely to report false positives.

Anyway, these are all details that may also have been overlooked by those in the field and that even if they realized it, they don't have a strategy (or see the need) to perform such meta-analysis. Another way that an outsider can bring freshness to a field.

Reply on news.ycombinator.com