[ImageJ-devel] imagej-ops : outputop-service branch

[Christian Dietz]

Hi Jay,

this is great feedback. Thank you. As you already mentioned in your
E-Mail, the API of the OpResolver, FeatureService etc is highly unstable
at the moment as the development still is on a topic-branch of
ImageJ-Ops (which is also in 0.x at the moment). However, I'm really
trying to get this stuff merged to master soon. Therefore, I will start
(and actually already started) a series of more atomar Pull-Requests on
ImageJ-Ops to get all the stuff merged. It would be amazing if you can
also have a look at the PRs and give some comments and feedback on the
individual functionality.

Please find more answers in-line. Again, thank you for your feedback!
Great that the stuff is used out there already ;-)

Christian



On Sa, 2015-06-13 at 03:29 -0500, Jay Warrick wrote: 
> Hi All,
> 
> 
> I chatted with Curtis Rueden a bit today about trying this
> developmental feature extraction stuff out. He said I should note any
> successes or challenges as it might help your progress. He suggested
> potentially to post on the pull request but it seemed a bit long for
> there.
> 
> 
> First of all, fantastic stuff. You are bringing so much quantitative
> power to the fingertips of a lot of individuals like myself. Thank
> you! Secondly, sorry if I'm using anything inappropriately in the
> rundown below but this is my best shot at trying to give you some
> feedback on my attempt to use your code. Also, I totally realize
> things are still in flux so I didn't necessarily have expectations
> that all this stuff would work immediately for me.
> 
> 
> As of what was synced to the imagej-ops : outputop-service branch
> tonight, I was able to successfully get measures from a
> `FirstOrderStatFeatureSet` for a set of `LabelRegion`s that I got from
> a `Img<UnsignedByteType>` using `ConnectedComponents` (from latest
> ImgLib2-algorithm SNAPSHOT but just temporarily copied to its own
> class local for this test) and apply them to an
> `Img<UnsignedShortType>` using `Regions.sample(myLabelRegion, myImg)`.

This is 100% correct. The new CCA should soon be part of the stable
imglib2-algorithms. We will also make it available in Ops as soon as it
was released with imglib2-algorithms. Thanks for tpietzsch for the
implementation!

Anyway, I'm working on several Ops will make your live easier when
trying to extract and loop over LabelRegions. 

> I think I did most of these tests using connected components detected
> using a 4-connected structuring element (if that matters at all).
> 
> 
> I ran into an issue with the `GeometricFeatureSet` where I got a 
> `````
> java.lang.ArrayIndexOutOfBoundsException: -1
> java.util.ArrayList.elementData(ArrayList.java:400)
> java.util.ArrayList.remove(ArrayList.java:477)
> net.imagej.ops.geometric.polygon.DefaultConvexHullOp.run(DefaultConvexHullOp.java:127)
> `````
> All my regions are quite sizeable (all between 1000 and 6000 pixels,
> see end of email) so I'm not sure why this would happen. I'm basically
> doing the following for this test using the same regions as I used
> with the `FirstOrderStatFeatureSet`
> 
> 
> `````
> GeometricFeatureSet opGeometric =
> ij.op().op(GeometricFeatureSet.class, LabelRegion.class);
> List<Pair<String, DoubleType>> results =
> opGeometric.getFeatureList(Regions.iterable(myLabelRegion));
> `````

The GeometricFeatureSet is really something we have to work on. Thanks
for the report. I also found some serious problems with the design of
several methods and I hope that I can fix it next week.

> 
> 
> I'm also having a problem with `Haralick2DFeatureSet` using those same
> `LabelRegion`s.
> `````
> java.lang.ArrayIndexOutOfBoundsException: -1
> net.imglib2.roi.labeling.LabelRegion.dimension(LabelRegion.java:459)
> net.imglib2.AbstractWrappedInterval.dimension(AbstractWrappedInterval.java:59)
> net.imagej.ops.features.haralick.helper.CooccurrenceMatrix2D.run(CooccurrenceMatrix2D.java:119)
> `````
> I'm testing the `Haralick2DFeatureSet` as follows...
> ````
> Haralick2DFeatureSet<UnsignedShortType> opHaralick2DHor =
> ij.op(Haralick2DFeatureSet.class,
> (IterableInterval<UnsignedShortType>) image, 8d, 1d, "HORIZONTAL");
> List<Pair<String, DoubleType>> results =
> opHaralick2DHor.getFeatureList(Regions.sample(myLabelRegion, image));
> `````

This is something we have to look into as soon as we file the Haralick
PR. I will turn your code into a test and make sure that it works.

> 
> The `HistogramFeatureSet` appears to work. I think I don't have my
> image adjusted right to see appropriate values in the histogram output
> though as I seem to get all 0's. I assumed the "number of bins"
> parameter would autoscale the data according to the image type (e.g.,
> UnsignedShortType image scaled to fit in 256 levels/bins) but I can
> see why that was likely a bad assumption.

This should actually happen. I will double-check. Again, as soon as we
file the "HistogramFeature" PR, it would be great if you could test
again. 
> 
> 
> The `ImageMomentsFeatureSet` seems to work fine.
> 
> 
> I got the following error trying to use the `Zernike` feature set...
> `````
> java.lang.NumberFormatException
> java.math.BigDecimal.<init>(BigDecimal.java:470)
> java.math.BigDecimal.<init>(BigDecimal.java:739)
> java.math.BigDecimal.valueOf(BigDecimal.java:1069)
> net.imagej.types.BigComplex.setReal(BigComplex.java:225)
> net.imagej.ops.features.zernike.ZernikeComputer.normalize(ZernikeComputer.java:193)
> net.imagej.ops.features.zernike.ZernikeComputer.fastZernikeComputation(ZernikeComputer.java:151)
> `````
> I'm using it as follows on the 8-bit mask image
> `````
> ZernikeFeatureSet<UnsignedByteType> opZernike =
> ij.op().op(ZernikeFeatureSet.class,
> (IterableInterval<UnsignedByteType>) mask, true, true, 1, 3);
> List<Pair<String, DoubleType>> results =
> opZernike.getFeatureList(Regions.sample(myLabelRegion, mask));
> `````

I will forward it to the implementor of Zernike. Seems that there are
some tests missing.

> 
> I figured, in case I'm doing a bunch of things wrong, it would be
> easier for you to just have the data I'm using to see if you get the
> same issues using your own code. The images that I'm using are here...
> 
> 
>  https://drive.google.com/folderview?id=0BzopfqjGpdu8flBNTHdjXzN2eUJhR1NCNmNTbHJ5T3RNVDJRbmdwVnJiYnI0Nk43ZXp3QzA&usp=sharing
> 
> 
> The folder includes a small txt output of only the label regions that
> I'm analyzing (their pixel areas and centers of mass), given I am not
> using all labeled objects that come from the connected components
> analysis. Lastly, my label indices start at 0 if you are looking to
> match up output from ConnectedComponents with this.
> 
> 
> If you have any feedback on where I might be doing things wrong,
> please let me know.
> 

Thanks for the data. We will also soon have a Image Segment Feature Node
within KNIME and also use the KNIME testing system to find the last(...)
bugs. Maybe this node can also help you extracting features for your
data!

> 
> Thanks!
> 
> 
> Jay
> 
> 
> 
> 
> _______________________________________________
> ImageJ-devel mailing list
> ImageJ-devel at imagej.net
> http://imagej.net/mailman/listinfo/imagej-devel