crazyseabiscuit / jfeaturelib Goto Github PK

Different feature descriptors are parameterized differently.

It would be nice to have a full Java based API for settings parameters.

Some examples:

PHOG has Setters for bins and recursions, but not for canny.

setGradientSource seems to be a bit odd. There is an initial value, but after 
the first run() it becomes null, making the feature descriptor somewhat 
non-reentrant.

AutoColorCorrelogram on the other hand apparently cannot be configured at all 
from Java, but only via properties, meaning you can't use different 
parameterizations in the same run if you want to compare them.

The Haralick implementation is missing Feature 14 (Maximal Correlation 
Coefficient).

See: 
Haralick, Shanmugam, Dinstein. Textural Features for Image Classification. 
1973. Page 10.

Catch Exception w/out logging or rethrowing in 
/trunk/src/de/lmu/dbs/jfeaturelib/ThreadWrapper.java
de.lmu.dbs.jfeaturelib.ThreadWrapper

        catch(InstantiationException | IllegalAccessException e){
            //FIXME Fix this or little kittens will die!
            System.out.println("Error during instantiation");
            e.printStackTrace();
        }

Log Exception and throw an InvalidStateException

provide simple tests for all features and - if possible also some sanity checks

PHOG sets the SUPPORTS_MASKING enum but in L.117ff there is no check accordig 
to the masks.

I dare say that PHOG actually dows NOT support masking.

Please, report which descriptors are able to work with binary masks. 
Currently, the only way to know whether a descriptor supports a mask or not is 
if you look at the source code.

A wiki page could be provided to describe which descriptors support masks and 
which don't. 

This issue was created by revision r230.

The detector should be refactored to work on graylevel images as it converts 
color images to grays in the code anyways

The shape descriptors only accept binarized images.

It should be thought about changing it to Supports.BINARY (which must be 
introduced)

Maybe it should be thought about pulling the automatic-conversion out of the 
extractors into a separate class.

It would be great if the Extractor utility could accept binary masks to extract 
image features. Maybe adding a new parameter to indicate the directory 
containing  corresponding masks to those images inside the image directory.

Best regards,

Patrícia.

This issue was created by revision r236.

This is a minor performance issue.
By using a 3x3 kernel, imageJ's methods can be used directly.

Can be replaced with a more efficient implementation

This issue was created by revision r308.

beautify/test cleanup Method Kernel.process()

currently the LibProperties can only be instanciated with a properties files.
This can lead to problems in some configuration.

A possibility to instanciate the LibProperties without file is appreciated 
(also for testing)

CUrrently the shape descriptors accept several kinds of images yet they really 
work only on binary images.

It should be checked if the application should be restricted to a ByteProcessor 
or BinaryProcessor

* http://rsbweb.nih.gov/ij/developer/api/ij/process/BinaryProcessor.html
* http://rsbweb.nih.gov/ij/developer/api/ij/process/ByteProcessor.html

        // mean values
        for (int i = 0; i < NUM_GRAY_VALUES; i++) {
            mu_x += i * p_x[i];
            mu_y += i * p_y[i];
        }

that doesn't compute the mean - and subsequently the variance is wrong as well

It has been reported that some shapre descriptors define the background as 0 
some as -1.
This should be unified to 0.

SIFT currently needs the binary path in the constructor.
This parameter could be passed via JVM-Env. variable or global properties

Hi,
as stated in the mail, please move the featurecomparator and compare package to 
your project as long as we don't have enough ppl to implement really a lot of 
comparison techniques and distance metrics.

It would be easier if the canny edge detector was added to PHOG extraction 
process. It could be called internally to do the PHOG extraction in only one 
step. 

Also, a validation check need to be done to accept only graylevel images. Or 
even better to convert color images to grayscale automatically, as it is done 
in PHOG algorithm (http://www.robots.ox.ac.uk/~vgg/research/caltech/phog.html)

This is in version 1.3.0

When I run PHOG on an image, the resulting feature is twice as large as it 
should be, with the first half of the features equal to zero.  

The reason for this is the following:  
  -- PHOG initializes the feature vector in the function 'void initFeature()'.  initFeature computes the size of the final size of the feature vector (correctly) and then allocates memory for it (feature = new double[length];).  These are all zero.

  -- In 'void buildHistogramRecursively()', the feature vector is _appended_ to by the histogram data  (feature = Arrays2.append(feature, histogram.getData());).  This adds features to the end of the vector (after re-sizing it).  

A solution is to change initFeature and set feature=new double[0].  
Arrays2.append handles this case, and results in the correct size feature vector

Attached are (yet untested) classes for RGB/HSB Histograms that allow non 
uniform splitting / sizing of the bins.

The content should/could be merged with 
de.lmu.ifi.dbs.jfeaturelib.features.Histogram

From an HowTo's comment:

When I add JFeaturelib as a Maven dependency in Netbeans (7.3), this happens: 
"Failed to read artifact descriptor for net.semanticmetadata:lire:jar:0.9.3: 
Could not transfer artifact net.semanticmetadata:lire:pom:0.9.3 from/to lib 
(file://${project.basedir}/lib): Repository path /lib does not exist, and 
cannot be created"

Is this some internal local dependency? I actually do have a "lib" folder, 
although it doesn't contain the lire pom/jar. I have lire available as a maven 
dependency, but JFeaturelib appears to be looking for it in a very specific 
place.


There should be a better solution for this.

There are some italian comments in harris corner detector 
http://code.google.com/p/jfeaturelib/source/browse/trunk/src/de/lmu/dbs/jfeature
lib/pointDetector/Harris.java

which should be translated to english

The Extractor class currently only supports FeatureDescriptor s that live in the

de.lmu.ifi.dbs.jfeaturelib.features

package (or below).

It should support third-party packages, too, as long as they implement the 
desired interface.

image class should not be required or at least use some reasonable defaults in 
commandline extractor.

Wither set a default-class or try to ignore the class if no class is set.

The AutoColorCorrelogram feature does not seem to work:

Exception in thread "main" java.lang.NullPointerException
    at net.semanticmetadata.lire.imageanalysis.AutoColorCorrelogram.extract(AutoColorCorrelogram.java:304)
    at net.semanticmetadata.lire.imageanalysis.AutoColorCorrelogram.extract(AutoColorCorrelogram.java:230)
    at de.lmu.ifi.dbs.jfeaturelib.features.AutoColorCorrelogram.run(AutoColorCorrelogram.java:62)

Maybe the bug is in lire, though.

This issue was created by revision r233.

The detector should be refactored to work on graylevel images as it converts 
color images to grays in the code anyways

Sometimes I've got phog bins with zero values. I don't know if it is an error 
in phog descriptor or not. 

I'm using chi-square distance to measure the distance between two phog feature 
vectors. Since this metric doesn't allow bins with zero values, it results with 
error.

I attach an image (28.jpg) which I got this behaviour (phog.txt, see 24th bin). 
Phog parameters are the deafults: bins=8 and recursions=1 

I used the Extractor utility to extract phog features. My command line: java 
-cp JfeatureLib.jar de.lmu.dbs.jfeaturelib.utils.Extractor -D PHOG -c c -d 
/home/patricia/imgs/ -o phog.txt

Best regards,
Patricia

Color Correlograms might be easy and wellknown to implement.

Image Indexing Using Color Correlograms

@inproceedings{huang1997image,
  title={Image indexing using color correlograms},
  author={Huang, J. and Kumar, S.R. and Mitra, M. and Zhu, W.J. and Zabih, R.},
  booktitle={Computer Vision and Pattern Recognition, 1997. Proceedings., 1997 IEEE Computer Society Conference on},
  pages={762--768},
  year={1997},
  organization={IEEE}
}

Please remove the "Lire"-Prefix in the classes 
LireCEDD
LireFCTH
LireFuzzy
LireGabor
LireJCD
LireTamura

Surafe Roughness Features

http://www.gcsca.net/IJ/SurfCharJ.html provides an ImageJ Plugin.

de.lmu.dbs.features.haralick.RoughnessCalculator might be an implementation. 
Possibly a wrapper around the plugin is more appropriate.

Scientific publication:
Chinga, G., Gregersen, O., Dougherty, B.,
"Paper surface characterisation by laser profilometry and image analysis",
Journal of Microscopy and Analysis, July 2003.