LabelMe FAQs

HOMEIMAGES = '/desired/path/to/Images';
HOMEANNOTATIONS = '/desired/path/to/Annotations';
folderlist = {'05june05_static_street_porter'};
LMinstall (folderlist, HOMEIMAGES, HOMEANNOTATIONS);

 HOMEANNOTATIONS = 'http://labelme.csail.mit.edu/Annotations';
HOMEIMAGES = 'http://labelme.csail.mit.edu/Images';
D = LMdatabase(HOMEANNOTATIONS, {'static_nature_web_outdoor_animal'});

[names, counts] = LMobjectnames(D);

HOMEANNOTATIONS = 'http://labelme.csail.mit.edu/Annotations';

HOMEIMAGES = 'http://labelme.csail.mit.edu/Images';

D = LMdatabase(HOMEANNOTATIONS, {'spatial_envelope_256x256_static_8outdoorcategories'});

 
for i = 1:length(D);

folderlist{i} = D(i).annotation.folder;

filelist{i} = D(i).annotation.filename;

end
 

LMthumbnailsbar(folderlist(1:500), filelist(1:500), 'page1.html', HOMEIMAGES);

LMthumbnailsbar(folderlist(501:1000), filelist(501:1000), 'page2.html', HOMEIMAGES);

This code will create two pages: page1.html and page2.html

First you have to create the index (this will take few minutes)

HOMEANNOTATIONS = 'http://labelme.csail.mit.edu/Annotations';

HOMEIMAGES = 'http://labelme.csail.mit.edu/Images';

D = LMdatabase(HOMEANNOTATIONS);

Then, you can create the list of images that you want using the function LMquery. If you only want pictures containing cars:

[Q,j] = LMquery(D, 'object.name', 'car');
clear folderlist filelist
for i = 1:length(Q);
      folderlist{i} = Q(i).annotation.folder;
      filelist{i} = Q(i).annotation.filename;
end

And now you can download those images onto your local machine with the LMinstall command:

HOMEIMAGES = '/desired/path/to/Images';
HOMEANNOTATIONS = '/desired/path/to/Annotations';
LMinstall (folderlist, filelist, HOMEIMAGES, HOMEANNOTATIONS);

Use the function LMobjectcrop, then loop on all images and objects:

for n = 1:length(D)

     if isfield(D(n).annotation, ‘object’)

          for m = 1:length(D(n).annotation.object)

               imgCrop = LMobjectcrop(img, D(n).annotation, m);

               % save imgCrop in some local folder

     end

   end

end

Use the function addsmallobjectlabel:

D = addsmallobjectlabel(D, height, width);

 

This function will add the label ‘smallobject’ to objects smaller than [height x width] pixels.

Then you can use LMquery to get rid of the small objects:

D = LMquery(D, ‘object.name’, ‘-smallobject’);

Many images are already fully annotated. In order to find those images, you can use the command LMlabeledarea. This function will return, for each image in the dataset, the percentage of pixels that are labeled:

relativearea = LMlabeledarea(D, objectname);

j = find(relativearea>.9);

This will return a list of images with at least 90% of the pixels labeled. This does not mean that all the objects in those images are annotated.

The scene description is stored in the field: D.annotation.scenedescription. So, to get all the kitchen images, you can do the query:

Dq = LMQuery(D, 'scenedescription', 'kitchen');
    

You can find a detailed example of using Wordnet with LabelMe in this Matlab script:
http://labelme.csail.mit.edu/LabelMeToolbox/demoWordnet.m
The toolbox does not require installation of Wordnet.

For the latest version of the annotation tool, we require a new directory (with write permissions) called "TmpAnnotations". In all, you should have the following directories:

Images/

Annotations/

TmpAnnotations/