CONTACT

   Brandon M. Smith
   bmsmith@cs.wisc.edu


CHANGE LOG

   Tuesday, May 22, 2013 - Version 1.0 released

   Thursday, Feb 18, 2016 - Corrected the segment labels for image 
      2371615952_1, both original and resized versions.

CITATIONS

   The citation for our segment label annotations, which supplement the Helen 
face dataset, is

   Brandon M. Smith, Li Zhang, Jonathan Brandt, Zhe Lin, and Jianchao Yang, 
   Exemplar-Based Face Parsing, IEEE Computer Society Conference on Computer 
   Vision and Pattern Recognition (CVPR), June, 2013. Project website:
   http://www.cs.wisc.edu/~lizhang/projects/face-parsing/

The citation for the original Helen face dataset is

   Vuong Le, Jonathan Brandt, Zhe Lin, Lubomir Boudev, and Thomas. S. Huang, 
   Interactive Facial Feature Localization, 12th European Conference on 
   Computer Vision (ECCV), 2012. Project website:
   http://www.ifp.illinois.edu/~vuongle2/helen/


SEGMENT LABEL FORMAT

   We provide 11 segment label types for each Helen image:

   Label 00: background
   Label 01: face skin (excluding ears and neck)
   Label 02: left eyebrow
   Label 03: right eyebrow
   Label 04: left eye
   Label 05: right eye
   Label 06: nose
   Label 07: upper lip
   Label 08: inner mouth
   Label 09: lower lip
   Label 10: hair

The segment labels for each image are stored in 11 separate label probability 
maps (LPMs), one for each label type, each saved as an 8-bit grayscale PNG 
image, where 0 indicates 0.0 probability and 255 indicates 1.0 probability. 
The file format for each LPM is:

   <Helen image name>_lblXX.png,

where <Helen image name> is something like "232194_1" and XX is the label 
number above. The LPMs for each Helen image are stored in the same directory; 
the directory name reflects the Helen image to which it corresponds. 


TESTING, TUNING, AND EXEMPLAR SETS

   The 100 testing images used in our paper (citation above) are listed in
testing.txt; the 230 tuning images are listed in tuning.txt; and the 2000 
exemplar images are listed in exemplars.txt. The format of each line is:

   <Helen image number> , <Helen image name>


IMAGE SIZES

   LPMs are provided in two different sizes: original and resized. "original" 
reflects the size of the Helen images. Only the LPMs for the original size are 
provided. The original Helen images and landmark annotations can be downloaded 
at 

   http://www.ifp.illinois.edu/~vuongle2/helen/

In the resized dataset, the faces are normalized so that the inter-occular 
distance (IOD) is approximately 55 pixels. All experiments in our paper 
(citation above) were done using these resized images. Additionally, the 
resized exemplar images have been rotated so that the eyes are approximately 
level; tuning and testing images were not rotated. The transformed images 
(./small/images) and landmark annotations (./small/points) are also provided 
with the LPMs (./small/labels) in the resized dataset. The transformed 
landmark annotations are given in the same format as the original Helen 
landmark annotations.


SEGMENT GENERATION PROCESS

   For Labels 02 through 09 (the internal face components) and the lower part 
of Label 01 (face skin) the segments were generated using the contours 
provided with Helen; these segments are generally very accurate and have 
sharp, well-defined boundaries. For Labels 00 and 10 (background and hair, 
respectively) and the upper part of Label 01 (face skin) the segments were 
generated using manually-annotated seed "scribbles" followed by an automatic 
matting algorithm (see our paper for more details); these segments generally 
have soft boundaries. The hair mattes are usually accurate, but mistakes are 
inevitable. Therefore, to ensure fair accuracy measurements, we manually 
annotated the face skin, hair, and background segments in all testing images.