MegaFace / Challenge

http://www.washington.edu/news/2016/06/23/how-well-do-facial-recognition-algorithms-cope-with-a-million-strangers/
http://megaface.cs.washington.edu/participate/challenge.html

Challenge 1: Recognition with varying number of distractors

For current results see the leaderboard!

Get Started

• Fill out this form to gain access to dataset. You will receive access information within 1-2 days if approved.
• View readme.txt included in development kit
• Please read the FAQ below.

Experiment

• Identification and Verification

  1. Download MegaFace and FaceScrub datasets and development kit
     
  2. Run your algorithm to produce features for both datasets
     
  3. Run our experiment script with 10, 100, 1000, 10000, 100000, 1000000 distractors
     
  4. Upload results into the google drive folder you received with access information. Please also upload links to features files for the full FaceScrub and MegaFace datasets
     
  More information about the experiment and the development kit files can be found in the development kit readme
  

Necessary Files

• Datasets

  • Our dataset (link)
  • FaceScrub dataset (by downloading this dataset you agree to the conditions of the creators of the data)
    • FaceScrub cropped + json zip — 12 GB
    • FaceScrub full tgz — 16.4 GB
    • FaceScrub bounding boxes actors txt
    • FaceScrub bounding boxes actresses txt

• Training Set

  • You may train with any set except for FaceScrub, MegaFace, and FGNET
  • Some systems are trained on millions of people, and others on several thousands. One of our goals is to compare face recognition algorithms independent of the training data. Thus please specify # of training photos and # of unique people you used for training. The results will be tiered accordingly, e.g., if you trained on 1K photos you won’t compete with groups that were trained on 1M photos.

• Linux Development Kit (.zip) (.tar.gz) — 16 MB

• Required Others

  • OpenCV (link)
    Open source computer vision and machine learning software library
    

Frequently Asked Questions

• What should we do if we cannot detect a face in some photos?

  • If you cannot detect a face in a photo then you should use our landmarks provided in the json files.
  • Landmarks meaning: landmark 0 is center of the right eye, 1-center of the left eye, 2-tip of the nose.
  • In case one or more of the landmarks is missing it means that the point is occluded in the photo.

• Why are there fewer FaceScrub feature files in your features than in the whole set?

  • We use a subset of FaceScrub for initial tests (to speed up the testing) but will use the full FaceScrub for additional tests, please compute features for the full set.

• Do we need to submit our features for the full Megaface and Facescrub?

  • Yes, please submit a link to all the megaface features and all facescrub features.

Please cite the paper if you use our code, results, or dataset in a publication (link)

View Current Challenge Results  

Optional Files

• Our Feature Files

  • MegaFace Features (.zip) (.tar.gz) — 5.4 GB
  • FaceScrub Features (.zip) (.tar.gz) — 183 MB