Options

Cluster Vision: Surface Temperatures 2010

Hello John and Hello to everyone

What if the human intuition and vision is too specific to cognize patterns in data, what if we used machine vision to identify 'keypoints' as features in temperature data (per day) and then use machine learning to cluster these keypoints (derivatives taken on the splined original data no normalization):

Cluster Vision: Laplacian Surface Temperatures 2010

Paul check out the Himalayas.

Cluster Vision: dtdx Surface Temperatures 2010

Mahler patterns are visible, so they are not artifacts of normalization.

Cluster Vision: dtdx Surface Temperatures 2010

In particular SURF keypoints, does not have to be SURF or SIFT we could choose any or make our own:

SIFT

SURF paper

SURF

Python SURF SIFT

Comments

  • Options

    In CDF form these videos could be stepped though day by day, you could also do that in Youtube. As you can see the images SURF keypoints are few about 20 or so, therefore we could collect them and their clusters, we might be able to use these points for extended periods of time to conduct classifications and weather patterns, as opposed to using the entire data sets.

    Unfortunately we could examine these sorts of ideas only after building prototypes and playing around with them.

    Dara

  • Options
    edited August 2014

    So SIFT/SURF features are designed to find gradient extremum patterns which are highly robust to (ie, almost invariant to) affine transformations (which are a big problem when trying to match points between real-world images). Is it obvious that this is a relevant criterion for use in data mining from statically co-ordinatized datasets?

  • Options

    So SIFT/SURF features are designed to find gradient extremum patterns which are highly robust to (ie, almost invariant to) affine transformations (which are a big problem when trying to match points between real-world images).

    YES! So FIRST you run the computer vision algorithm get these points (or regions) and SECOND you run on the latter the machine learning algorithms e.g. clustering classification and FINALLY forecast. The advantages are that the points selected are fewer and most relevant.

    Is it obvious that this is a relevant criterion for use in data mining from statically co-ordinatized datasets?

    Do not know what you mean 100% but I see either this method or brute force use all the pixels in the data

  • Options

    These are the computer vision algorithms which runs fairly fast in Mathematica:

    Computer Vision in Mathematica

    Obviously we could add our own if needed.

    SO you could get an idea

    D

  • Options

    So the point I was making is that SIFT/SURF are designed to be "robust to affine changes" rather than "interesting" -- SIFT/SURF will throw out patterns that are more interesting in favour of ones that should be more robust. It's not clear that this robustness is useful. It might be that going back to older computer vision interest points which aren't as useful for actual computer vision -- because they're not robust to camera motion -- for the purpose of doing interest point detection in statically gridded data. (If there was some reason to believe affine transformations were going on in the data SIFT/SURF might be good.)

  • Options

    So the point I was making is that SIFT/SURF are designed to be “robust to affine changes” rather than “interesting” – SIFT/SURF will throw out patterns that are more interesting in favour of ones that should be more robust.

    I could not say that. After all what is INTERESTING? Someone looking at 150MB of volumetric data and finds a point in the middle and feel something interesting about it?

    Again I have no idea what ROBUST means.

    You see we have downloaded chuck full of GPM satellite data on rain and wind and temperature and so on, one orbital of data is HUGE! and all of it look to me interesting and robust.

    The classic mistake of all machine learning developers is that they have a priori concept of something of interest or how things should come out in forecast! The idea is to let the data and code show us the way, than preconcieved notions.

    I won't push you to think this way or that way, that is your prerogative. But I am faced with terrabytes of data per hour in a few days from one group of satellites.

    Dara

    D

  • Options
    edited August 2014

    By "robust to affine changes" I mean the goal of the SURF interest point operator is to find a patch P in an image such that

    SURF(P) approximately equals SURF(A(P))

    (same for SIFT) where A is an affine mapping. This is a good design goal for image processing because local patches in the 3-D world undergo local affine transformations in the camera images as the camera moves. There's a lot of "interesting" structures in images that SURF decides to throw away because it can't make it fit the equation above. In other words, the points SURF gives you may well be interesting, but there are also going to be lots of interesting points it won't give you.

    I was just suggesting that it might be useful to try older "interest point operators" that don't include affine invariance in their requirements.

  • Options

    There’s a lot of “interesting” structures in images that SURF decides to throw away because it can’t make it fit the equation above.

    Could you give some examples? please.

    Paul's sloshing concepts or wind/water streams are robust, for that matter anything else discussed in this group, since they are affine transformation independent. Sloshing is sloshing no matter how you rotate or scale the data set.

  • Options

    I'll try to post a detailed respond soon. (I'm working on something else at the moment.)

  • Options

    Take your time, this takes a few years of back breaking work to bring to fruition, not a quick thingy

Sign In or Register to comment.