Face Recognition dataset

IMDb-Face dataset

• IMDb

  • Available at https://github.com/fwang91/IMDb-Face (IMDb-Face.csv)

• For fast processing, multi processing is supported with concurrent.futures module.

Running instruction

• Download 'IMDb-Face.csv' file from https://drive.google.com/open?id=134kOnRcJgHZ2eREu8QRi99qj996Ap_ML
• python imdb_crawl.py
  • Arguments

  -c: whether you crop the image with bounding box
  -d: delete existing data directory be
  

• If you save non-cropped image, corresponding bounding box will also be recorded in bb.txt file for each direcory.
• Make sure 'IMDb-Face.csv' and 'imdb_crawl.py' are located in same directory.

Megaface dataset

To run megaface test including identification(1m distractors), verification(@1e-6),

1. Download distractors and probe dataset
2. Preprocess dataset
3. Generate bin files with a trained face recognition model
4. Run megaface devkit

Download dataset

• Distractors
  wget -c --user 'id' --password 'pwd' http://megaface.cs.washington.edu/dataset/download/content/MegaFace_dataset.tar.gz
• Facescrub
  
  • wget -c --user 'id' --password 'pwd' http://megaface.cs.washington.edu/dataset/download/content/downloaded.tgz
  • python facescrub.py
    But, many urls can not be accessed.
• Both datasets can be accessed at http://megaface.cs.washington.edu/participate/challenge.html
• Dataset structure

   MEGAFACE -- distractors -- parent id -- ids -- images
            |                                  |- json file for each image 
            |
            |- facescrub -- ids -- images, bb.txt
                                |- bb.txt
   

• Facescrub bounding box file
  • facescrub.py
  • Need facescrub bounding box actor.txt, facescrub bounding box actress.txt files at megaface
  • Arguments

    - txt_files: [facescrub_actor.txt, facescrub_actress.txt]
    - timeout: timeout seconds for accessing url (need for downloading image)
    

    • This script file can also download image from text files, but many rows have missing urls.

Preprocess

Preprocess with your face detection/alignment model.

Generate bin files

• gen_megaface.py (Need your face recognition model)

  • Make bin files of megaface distractors/facescrub images from trained face recognition model.

  • Arguments

    - megaface_path: path of pre-processed distractor images
    - facescrub_path: path of pre-processed facescrub images
    - megaface_noise: noise list of distractors
    - facescrub_noise: noise list of facescrub
    - megaface_bin_path: distractor bins save directory
    - facescrub_bin_path: facescrub bins save directory
    - ckpt: trained face recognition model
    - file_ending: file ending name, ex) _baseline.bin: aaa.jpg -> aaa_baseline.bin    
    

  • Resulting bin files of gen_megaface.py (ex: file_ending: _baseline.bin_)

    _baseline -- facescrub_bin -- ids -- bin files (***_baseline.bin)
                             | 
                             |- megaface_bin -- parend id -- ids -- bin files (***_baseline.bin)
    

Run megaface devkit

• On terminal,
  python run_experiment.py --file_ending _baseline.bin --out_root baseline_results -d

  • Need at least 32G memory
  • Need devkit from http://megaface.cs.washington.edu/participate/challenge.html
    • bin/FuseResults, Identification
    • models/jb_identity.bin, jb_LBP.bin
    • templatelists

• run_experiment.py

  • Executes identification and verification binary files (bin/Identification, bin/FuseResults).
  • Arguments

    - distractor_feature_path: distractor bin files path (megaface_bin_path of gen_megaface.py)
    - probe_feature_path: facescrub bin files path (facescrub_bin_path of gen_megaface.py)
    - file_ending: file ending format (file_ending of gen_megaface.py)
    - sizes: number of distractors, set as [1000000] 
    

• Caution

  • binary files (bin/Identification, bin/FuseResults) can be only executed on opencv2.4.

Reference

• Megaface devkit
  • Can download at http://megaface.cs.washington.edu/participate/challenge.html
• distractor_noise_list, facescrub_noise_list
  • Megaface dataset has some noisy dataset (some images in facescrub are included distractors)
    • arcface
    • arcface github