SVM and GIST example please?

[Dave Mobley]

Last post was actually by me (but forgot to log in-not sure iof I can fix it after the fact)

Hi All,

I have a data set (some thousands of images). I want to run the GIST global descriptor on then and then apply SVM to it.

My basic procedure is:
1) Load dataset (Images are on filesystem in 12 distinct folders)
2) Run SVM (and feature extractor is a GIST extractor I made)

Does this sound like a correct approach? And if so, here is my code so far - any recommendations?

public class GistExtractor implements FeatureExtractor<floatfv, fimage="">
{
private Gist<fimage> gist = null;</fimage></floatfv,>

public GistExtractor(Gist<FImage> gist)
{
    this.gist = gist;
}

@Override
public FloatFV extractFeature(FImage object)
{
    gist.analyseImage(object);
    FloatFV fv = gist.getResponse();
    return fv;
}

}

public class GistCalculator
{
public static Gist<fimage> fsg = new Gist<fimage>(256, 256);
public static FeatureExtractor<floatfv, fimage=""> gex = new GistExtractor(fsg);
public static SVMAnnotator svm = new SVMAnnotator(gex);</floatfv,></fimage></fimage>

public static void main(String[] args) throws IOException
{       
    // load Dataset //
    GroupedDataset<String, VFSListDataset<Record<FImage>>, Record<FImage>> allData = 
            ScenicOrNot.getData(ImageUtilities.FIMAGE_READER);

    // group dataset //
    GroupedRandomSplitter<String, Record<FImage>> splits = new GroupedRandomSplitter<String, Record<FImage>>(allData, 10, 0, 2);

    // run svm //
    svm.train(splits.getTrainingDataset());
}

}

I get an error "Data did not contain exactly 2 classes. It had 12."

Does anyone have any samples that might use GIST + SVM? Currently my 12 folders are just that, folders. The names hold no relevance to what is inside (I actually use a RecordReader and extract a float score from a csv file for each image).

Thanks
-scprotz

[Jonathon Hare]

Your overall approach looks fine - I think the problem lies with your ScenicOrNot class, which I would guess is building a dataset with too many classes (for a binary problem the GroupedDataset should have two just groups).