• Nov. 26, 2013: Workshop schedule posted.
Nov. 25, 2013: Results released.
• Oct. 4, 2013: Minor update to the devkit to fix one of the test bounding boxes.
• Sep. 22, 2013: Test data released, registration and submission server available.
Jul. 5, 2013: Development kit and training data have been released.
Mar. 19, 2013: Fine-Grained Challenge 2013 will be run in parallel with ILSVRC2013
• Mar. 4, 2013: Organization is underway for the fine-grained challenge.

Challenge Details
This challenge will target fine-grained classification, i.e. classification among categories which are are both visually and semantically very similar. This is a very difficult regime which is even challenging for humans without careful training, and is critical for establishing a more detailed understanding of the visual world.

Specifically, this challenge will span a range of domains, and within each domain will test classification among the many fine-grained categories. The final score is the average accuracy, where the average is taken over the accuracies for the different domains. The winners will be invited to give a presentation during the FGComp portion of the ImageNet challenge workshop at ICCV.

There are two tracks.

Track 1:
In Track 1, bounding boxes will be provided at both training and test time. In addition, at test time the domain (aircraft, bird, car, dog, shoe) will be provided. Overall score is determined by the mean accuracy, taken across domains.

Track 2:
In Track 2, bounding boxes will be provided only at training time. Although the datasets for the different domains are not detection datasets, in practice, ground truth bounding boxes are rarely available at test time. As in Track 1, at test time the domain will be provided, and overall score is determined by the mean accuracy, taken across domains.

Data for the challenge is drawn from a number of fine-grained datasets, many of which have not been previously made available. The challenge includes many categories of aircraft, birds, cars, dogs, and shoes. Each training image is annotated with a ground truth bounding box around the object of interest.

To download training data, see the instructions included in the development kit.

To download testing data, please register. Instructions for downloading test data will be sent to you.

To register, please fill out this form. Upon doing so, an email will be sent to you with instructions for downloading the test data and making submissions.

Development Kit
The development kit may be downloaded here. It includes instructions on how to download training data, details of the annotation format, and the format for entry submission.

Jul. 5: Dev. kit, training data release
Sep. 22: Test data release, registration and submission server online.
Nov. 14, 5 pm Pacific time: Deadline for submissions.
Dec. 7: Presentation of results and winner's presentations at the FGComp portion of the ImageNet challenge ICCV workshop.

Alex Berg (UNC Chapel Hill)
Ryan Farrell (Brigham Young)
Jonathan Krause (Stanford)
Fei-Fei Li (Stanford)
Jia Li (Yahoo! Research)
Subhransu Maji (TTI Chicago)

Special Thanks
We thank the following people for their helpful comments and suggestions:    
Serge Belongie
Trevor Darrell
Andrew Zisserman
The organizers can be reached at

FGComp2013 is primarily self-funded, though we thank the organizers and sponsors of The Second Workshop on Fine-Grained Visual Categorization for contributing prizes.