LabelMe: A web-based tool for sharing annotated images

Research on object and scene recognition requires large amounts of real world images with ground truth. Those labeled datasets would require a big effort to build by a single group. The goal of the LabelMe database is to provide researchers with an online annotation tool so that everybody can help to build such a database and instantly share such annotations with the community. LabelMe is not only a database of labeled images, but also a web-based tool for easily creating and sharing more labeled data.

The LabelMe database contains thousands of labeled images and is growing every day. To get the database, we only ask you to contribute to it by annotating few images. After labeling some images, you will be granted access to the full database, and that will include your latest annotations too and a MATLAB toolbox for manipulating the database.

The labeling tool, plus a set of more than 10,000 images (4000 of which have been labeled), can be found at www.csail.mit.edu/~brussell/research/LabelMe/intro.html