CONTACT

Brandon M. Smith
Computer Sciences
1210 West Dayton Street
Madison, Wisconsin USA
http://www.cs.wisc.edu/~bmsmith
bmsmith@cs.wisc.edu


VERSION

September 11, 2014


LICENSE 

The contents of this download, including our supplementary landmarks, are 
released under the Creative Commons Attribution 4.0 International Public 
License, available with this download in license.html or online:
http://creativecommons.org/licenses/by/4.0/legalcode


CITATIONS

If you use our supplementary landmarks, please cite the following two papers 
(our paper and the original AFLW paper):

   Brandon M. Smith and Li Zhang, Collaborative Facial Landmark Localization 
   for Transferring Annotations Across Datasets, In European Conference on 
   Computer Vision (ECCV), September 6 - 12, 2014
   
   Martin Koestinger, Paul Wohlhart, Peter M. Roth, and Horst Bischof, 
   Annotated Facial Landmarks in the Wild: A Large-scale, Real-world Database 
   for Facial Landmark Localization, In First IEEE International Workshop on 
   Benchmarking Facial Image Analysis Technologies, 2011.


INTRODUCTION

We provide 64 supplementary landmark types and we fill in missing landmarks 
among the 21 types that AFLW defines for a total of 85 landmark types. The 
landmark definitions are given in landmark_definitions.pdf. Note that 
landmarks 1 - 21 correspond to the original AFLW types, and landmarks 22 - 85 
correspond to our supplementary types. 

Our supplementary landmarks were generated using the 2012-11-28 version of the 
AFLW landmarks as constraints. Because "some [AFLW] annotations still lack in 
quality" according to the AFLW documentation, and because some of the AFLW 
landmark definitions are subtly different than the definitions in the source 
datasets, we allow for some spatial adjustment among the original 21 AFLW 
landmarks. Therefore, landmarks 1 - 21 may not be exactly the same as those 
provided by AFLW. The images and the original landmark annotations are 
available from the AFLW website: http://lrs.icg.tugraz.at/research/aflw/

Many of the AFLW faces have very extreme head pose (e.g., +/- 60 degrees yaw 
or pitch rotation) and/or very unusual facial expressions that are beyond the 
range of head poses and facial expressions commonly available with landmark 
annotations in our source datasets (CMU Multi-PIE, Helen, LFPW, AFW, and 
IBUG). Our results on these faces are often poor. For example, on full-profile 
faces, our method will hallucinate landmarks for the occluded far side of the 
face; the hallucinated landmarks will tend to either extend away from the face 
or get jumbled within the face region. Therefore, in order to provide useful 
supplementary landmark annotations, we have manually removed these poor 
results. We provide supplementary landmark annotations for the 16473 remaining 
(passable) faces. Please note that the remaining landmark annotations are not 
always perfect, however. 


DATA FORMATS

We provide two formats: text files and MATLAB. The text files are stored in 
the "landmarks" directory. Each text file is named as 

   <5-character image file name>_<2-character face number>.txt. 
   
For example, landmarks/0/01565_02.txt provides landmarks for the the second 
face in the image flickr/0/01565.jpg. Each text file has the following format:
   
   <result quality label, currently always 0>
   x1 y1
   x2 y2
   x3 y3
   ... 
   x85 y85
   
where landmarks 1 - 21 correspond to the original AFLW types and landmarks 
22 - 85 correspond to our supplementary types. 

In the MATLAB file landmarks.mat, the data are stored as follows:

   data{i}.filename     % string
   data{i}.points       % 85 x 2 x F
   data{i}.quality,     % 1 x F

where i is the number in the AFLW image filename ranging from 1 to 71361. 
data{i} is empty if the AFLW image does not exist or if we do not provide 
supplementary landmarks. data{i}.filename is the AFLW image file name. 
data{i}.points is an 85 x 2 x F matrix of 85 landmark point locations, where 
F is the number of faces in the image for which we provide supplementary 
landmark annotations. data{i}.quality is 1 x F. data{i}.quality is currently 
always 0.