Thanks for your question and feedback.

We give general labeling guidelines here:

http://labelme.csail.mit.edu/guidelines.html

In general, we prefer that you label objects as if they were not occluded, whenever possible. When you do this, you can reliably infer the depth-ordering of the objects (c.f. one way of doing this is outlined in our IJCV paper: http://people.csail.mit.edu/brussell/resea...IM-2005-025-new.pdf).

Edited by author 01-31-2008 01:46 PM
Partialy visible object labeling

This question about labeling probably has been answered already, but you may want to post clear instruction how to draw the lines around the object if it is only partially visible (eg. chair only legs are seen, or plate or car when only quarter is seen, there is a person in front of the car, or only half of fridge is seen in the kitchen - creates triangular fridge). This creates weird geometric shapes and may may not really help you all that much. Garbage in garbage out, you know the old saying.... But other than that, keep up the great job.

Q: Should you label only what is visible even if it is partially occluded, or should you label the object as it should be seen.

Thanks! This is great work!

Yes, you may query using the Matlab toolbox:

HOMEANNOTATIONS = 'http://labelme.csail.mit.edu/Annotations';
D = LMquery(LMdatabase(HOMEANNOTATIONS), 'object.name', 'road');

The structure "D" will have the list of folders and images containing at least one car.

Hi! Is there any way we can obtain a list of images that contain an object of interest. For example, a query which gives me the list of all images on the database that contain at least one car.

Hi! Nice site!

Thanks for contributing labels and for your suggestions. In response to your suggestions:

(1) We are close to having Safari support but have not completed this yet. If you urgently need Safari support, then we encourage you to download the annotation tool source code: http://labelme.csail.mit.edu/Documentation/sourceCode.html

(2,3) Quality control in LabelMe is certainly an important issue. However, we have found that users generally provide good labels (c.f. our paper on the LabelMe database: http://people.csail.mit.edu/brussell/resea...IM-2005-025-new.pdf).

I just labeled one image. It's a nice interface. Suggestions: (1) support for Safari (Mac OS X), (2) you could also just show an image + existing contours (without labels) to the user (either created by previous users or computed via a segmentation algorithm) and ask them to either reject the contour or provide their own labels for it. (3) you could do the reverse too: show them just the image + a label (obtained from a previous user), and ask them to either state that no such object exists in the image, or contour that object... at the end, you can score the contour using an edge detector (or similar)... this way, you could perhaps improve/reject the badly drawn contours, and get rid of "maliciously" drawn objects which don't actually exist in the image.

Nevermind, I found the demoWordnet.m file.
This is great!

I can't seem to figure out how to use MatLab toolbox to be able to query the images using WordNet. Any help would be appreciated. Thanks!

The answers for your questions can be found via links from the main LabelMe page (http://labelme.csail.mit.edu). More specifically, check out the (1) tool source code and (2) search objects/scenes links for answers regarding source code and object queries.

Hey , i wanted to some details abt ur labelme..
Firstly, wich language are u ppl using ....
Secondly, How can i query
Thirdly do we have to label eac image for querying.... ??

Hello! great idea of color of this siyte!