Automatic analisis.

Discuss your experiences with and ideas about Stardust@home here.

Moderators: Stardust@home Team, DustMods

jasong
Posts: 2
Joined: Fri Jul 07, 2006 10:41 am

Maybe someday this will be a true DC project

Post by jasong »

Maybe someday, this will be a true Distributed Computing project, rather than a Distributed Scanning project.

Maybe when we actually discover the dust, someone will be able to come up with a computer program that can be trusted to find the stardust on it's own. Meanwhile, it's still a very cool concept.

Hawkwings
Posts: 12
Joined: Mon Jul 31, 2006 11:36 pm
Location: Oregon, USA

Post by Hawkwings »

The problem is that computers are bad at this sort of thing. And we'll probably finish before such a computer is made.

You know those security things when you sign up for something? Telling you to "Enter this security code here" and it's a bunch of numbers and letters that are partially obscured? Yeah, computers can't even do that yet. This is much harder, and much farther off.

TaeBoX
Posts: 3
Joined: Tue Aug 01, 2006 2:59 am
Location: South Dakota

Post by TaeBoX »

I don't know about you guys, but there is something much more exciting to me about finding the particle myself. It's much less rewarding to run a DC client and just hope you are helping.

Wolter
DustMod
Posts: 457
Joined: Mon May 22, 2006 2:23 am
Location: Enkhuizen, the Netherlands

Post by Wolter »

My thoughts exactly TeaBox.

Although i think the majority here are also on one of the DC projects, this project is truly about volunteer work. Spending hours at your PC to find the tiniest bits of dust in the hope they will give us a better understanding on how we became to be here. It's a great project, bringing a lot of us together. Let's make the best of it.
Just dusting... Image

Gyzuh
Posts: 2
Joined: Thu Aug 03, 2006 9:35 am

Automated detection

Post by Gyzuh »

Hi stardust enthousiasts,

There is one thing that I don't understand about this whole project. Why have these scientists immediately chosen for this manual approach? I agree that the development of a perfect detection program within a few months or years will be far too costly if not impossible, but I see still good uses for a computer as a filter to find the focus movies which most likely contain a particle. I agree that many other interesting phenomena and a few particles might be missed, but I think that this approach gives much sooner results and manual analysis can still be done anyway. To make the chance of finding a particle even bigger, the project leaders could have created another project in which people can write their own detection programs and test them on all focus movies. This requirese of course that all focus movies are available to the public. The result of a detection program should be a standardized list of candidates in order of likeliness, which can be fed to the virtual microscope by the same person for further analysis.

By the way, I have a bachelor of electrical engineering. I did an internship on my university on face recognition. That means that I had to do research on how to find a face in a picture. I have therefore some experience with techniques for finding objects in a picture. Having seen a couple of test movies, I think that similar techniques might well be applied on these focus movies. I would like to encourage the project leaders to publish their attempts to automate this process or reasons why this cannot be done on the website.

Regards,
Hans Giesen

jlczuk
Posts: 5
Joined: Wed Aug 02, 2006 10:10 am
Location: State of New York
Contact:

Post by jlczuk »

Public relations my friend, public relations! Besides, when it comes down to it, a trained eye (with the brain behind it of course), is a far better filter than anything a computer can acheive, albeit a bit slower.

WeBeGood
Posts: 65
Joined: Thu Aug 03, 2006 7:26 am
Location: Texas, USA

Post by WeBeGood »

jlczuk wrote:Public relations my friend, public relations! Besides, when it comes down to it, a trained eye (with the brain behind it of course), is a far better filter than anything a computer can acheive, albeit a bit slower.
I agree, public relations and it's much more fun and educational. To me it looks like it should be simple to automate with computers. Every camera you buy these days can auto focus, so it would be simple to scan the focus and look for things that are in focus. When everything is in focus, that's the surface. When only one thing is in focus, it's something else.

Sasari
Posts: 44
Joined: Mon Jul 31, 2006 3:17 pm

Post by Sasari »

I'd say recognizing a face is 100s times easier than finding a track.

With a face there are so many things you already know and can calculate. Eye posistion relative to nose etc. I would guess you would look for many things and how they are relative to other features. Determine angles on other features etc.


With face recognition you don't have to look at it as a movie, you look at a still shot. With tracks you can't look at a still shot and make any assumptions. You have to look at it as a movie. There are many single frame, even 3-4 fram sequences that show results that could be a track, but then in a 20 frame movie you can tell its not a track.

I would say that even determining the degree of focus in a movie alone would be pretty challenging. When the camera is above the gel, below the gel. If the entire movie is focusing at the same time.

Once the focus is determined the first thing I would look for is dark pixels appearing when there use to be no pixels.

It would be far less challenging if the tracks were the big easy sample tracks we get. However the tracks we are looking for I believe are like this

http://stardustathome.ssl.berkeley.edu/ ... e_id=41998

I've done image recognition on graphic images before. If I was looking for something that was 20x20 pixels in size it wasn't too bad on processor power. But when I started doing stuff in the 4x4 size like would be needed for the track in example I showed it started to be a major CPU hog. And this was image recognition on a still single frame. I can't imagine the processing power that would be needed to do this on a live movie.

WeBeGood
Posts: 65
Joined: Thu Aug 03, 2006 7:26 am
Location: Texas, USA

Post by WeBeGood »

If you look at your example, move the mouse up and down fast, between the surface (everything in focus) and the bottom (max depth). You'll notice surface features grow in size and get dimmer, while the suspected track does the opposite and shrinks towards a point and gets darker. The sample (MIR) movies shrink down to a nice sharp dark circle.

JohnPenny
Posts: 21
Joined: Tue Aug 01, 2006 3:55 pm

Post by JohnPenny »

Sasari wrote:I'd say recognizing a face is 100s times easier than finding a track.

With a face there are so many things you already know and can calculate. Eye posistion relative to nose etc. I would guess you would look for many things and how they are relative to other features. Determine angles on other features etc.

You could equate finding the nose and eyes in the image to finding 3 tracks. Then there is the calculation and comparison to a database....

With tracks you have a fairly uniform set of rules. I honestly dont think it would be that hard.

I do however think that it would be bad to automate entirely, because a human can recognise that a movie is out of focus instantly, and sometimes spot a track instantly, where as a programe wouldnt, and waste time.

jsellers
Posts: 10
Joined: Wed Aug 02, 2006 9:32 pm

Post by jsellers »

Sasari, nice find. I'm no expert, but it definitely looks like you've got something embedded in the gel. It could be that the track is just not there to visibly lead to the surface. I also suspect that the calibration movies might be throwing people off. They probably serve best to reveal who just isn't even trying, but as the tutorial said, we should all be keeping an open mind to what these tracks look like, because no one really knows. And if there is something that becomes clearer as the microscope focuses deeper below the surface, than why not assume it could be a track or a particle? I'd like an administrator to correct me if I'm misleading people, but I'd say click on something like that (even if it is just a lump in the aerogel batter!)

Gyzuh
Posts: 2
Joined: Thu Aug 03, 2006 9:35 am

Post by Gyzuh »

I have some comments on your replies, but first I want to say that the face recognition technique that I referred to is the technique that is described in a paper called "Feature extraction from faces using deformable templates" by Alan L. Yuille, David S. Cohen and Peter W. Hallinan.

jlczuk: I think that it's indeed all about public relations. I also think that they actually hope to discover some other unknown phenomena, which is impossible with a computer. I think that they are not even that much interested in the dust, because they probably already have a pretty good prediction of the composition of the dust, which just needs to be confirmed.
By the way, I did not say that the computer is better than the human eye. I just wanted to point out that the computer would be a useful addition to the process, because it can do an early selection of most likely candidates. I guess that a computer could for example filter out the movies that are not in focus. Even if it works only in 90% of the cases, it would still be useful, because it lowers the time that is needed to find the first track, although the time to find all tracks with 100% certainty will not be reduced.

Sasari: I don't agree that recognizing a face is probably 100 times easier.
First of all, faces have pretty irregular shapes, which differ between people. I assume that the tracks have a nearly circular shape and a size comparable to the examples. There is probably no proof that this is always the case, but I'd say that it's very likely that there will be at least 1 similar track among the 8 tracks that are expected. The advantage of a circular shape is that it can easily be described by a mathematical formula, which makes the use of a database superfluous. Even if we'd use a database, I'd say that it will be only a very small database.
Second, you are right that we are dealing with a series of images instead of a single image. However, many image recognition algorithms deal also with several different representations (with for example intensity and edges) of a single image. I expect that you can already get reasonable results with just an edge representation of every image in the focus movie, while you need to extract more representations for face recognition, so it is not as worse as you say. Moreover, the search for the track itself needs only to be done on images below the surface, so once the surface has been found, a part of the movie need not be searched anymore.
Third, assuming that the whole area is in focus in one image of the movie, I think you could determine the focus by making an edge representation of each image, summing all pixels of each edge representation together and determining the image with the highest total. The whole area is probably in focus in one image anyway, unless the movie was taken close to the edge of a piece of aerogel, which can be determined easily. I assume also that neighbouring focus movies have their focus at the almost the same height, so we don't need to check the whole movie. Besides, I don't think that it really matters if the focus is a little bit lower or higher than the focus that was found.
Concludingly, I think that the computing power needed for a focus movie is much lower than the computing power needed for face recognition. Anyway, I think that NASA must have some supercomputers available for this kind of research, so they have a lot more computing power available then we have.

JohnPenny: I think that the only reason why we should not automate this process entirely is, because we cannot obtain 100% accuracy. I think that any human needs a couple of seconds to move between the images and find the focus. A computer can certainly find a focus with for example 95% accuracy within a tenth of a second.

I think the most important thing to notice is that anything that is out of focus has edges that are less steep than objects that are in focus. If we find some steep edges below the surface, while they disappear when move up to the surface, we have probably found a track. That seems at least to be the case in the sample movies.

I know that I'm making some assumptions, but that is no problem, because I just wanted to convince you that it is possible to find the first tracks faster with the help of a computer, not to obtain a 100% coverage of all tracks.

maurix
Posts: 4
Joined: Sat Aug 05, 2006 2:10 pm

Automatic analisis.

Post by maurix »

I saw the images and i think that can be possible a parametrization of each focus movie so that all the data can be reduces and statistically analized in automatic form.
I undrstand that a person can do the work better and i like this comunity system but, are you guys using an automatic system too? A computer can analize 10millon images in only some days!
I was plaing arround with fft of block of 32x32 of images on test images and i think the algoritm can reach easly a sensitivity of 100%, detect bad focus and reject tons of images!
I will follow with personal inspection for fun but i hope you can give us more details of analisis process. Tnks

icestation
Posts: 13
Joined: Sat Aug 05, 2006 9:51 am

Post by icestation »

i think that computer can not distinguish real traks , its even hard for humans . taught we dont even know precisly how real traks seems like. so i dont think it can be an algorytm for that .

icebike

Post by icebike »

icestation wrote:i think that computer can not distinguish real traks , its even hard for humans . taught we dont even know precisly how real traks seems like. so i dont think it can be an algorytm for that .
Sure they can. You should see what computers can do with CAT Scans. They can build 3dmodels of the inside of your scull.

But it will take two years to fund, write, debug, and test that software, and in the meantime 10 thousand pairs of eyes are finding things faster.

And having more fun.

If this was going to be the first of hundreds of aerogel missions computer analysis would be the way to go, but its quite probably a one time event, and this is a briliant use of the internt and volunteers.

Post Reply