Labeled Face Parts in the Wild (LFPW)
Dataset
Kriegman-Belhumeur Vision Technologies, LLC.
Release 1 of LFPW consists of 1,432 faces from images downloaded
from the web using simple text queries on sites such as google.com, flickr.com,
and yahoo.com. Each image
was labeled by three MTurk workers, and 29 fiducial points, shown below, are
included in dataset. LFPW was
originally described in the following publication:
"Localizing Parts of
Faces Using a Consensus of Exemplars,"
Peter N. Belhumeur, David W. Jacobs, David J. Kriegman, Neeraj Kumar,
Proceedings of the 24th IEEE
Conference on Computer Vision and Pattern Recognition (CVPR),
June 2011.
[pdf]
[poster][project page]
LFPW was used to evaluate a face part (facial fiducial point)
detection method which was trained on 1,132 images and tested on 300 images.,
and so the data set is divided into two files, one for training and testing.
The database can be downloaded from here:
Due to copyright issues, we cannot distribute image
files in any format to anyone. Instead, we have made available a list of image
URLs where you can download the images yourself. We realize that this makes it
impossible to exactly compare numbers, as image links will slowly disappear
over time, but we have no other option. This seems to be the way other large
web-based databases seem to be evolving.
Data
Format
The
dataset is in a CSV file, using tab characters ('\t') as the delimiter. The
first line of the file defines the fields. The first is "imgurl",
which is the original url of the image. The second is "worker", which
can be one of "worker_0", "worker_1", "worker_2",
or "average". These are different workers' labels, along with an
average we computed (arithmetic mean for x, y locations, and median for type).
Most images thus have 4 lines in the csv file, one corresponding to each worker
and one for the average. (Note that there were a few errors with some of the
workers, so some images only have 3 rows.)
Following
these are all the fiducials. Each fiducial has 3 fields: <fid>_x,
<fid>_y, <fid>_t, which correspond to the x and y locations (with
subpixel accuracy) and the visibility type. The visibility can take the
following values:
0: Visible
1:
obscured by hair/glasses/etc.
2: hidden
because of viewing angle
3: hidden
because of image crop
Here are
the different fiducials (35 in all):
Citation
The database
is made available only for non-commercial use. If you use this dataset, please
cite the following paper:
"Localizing Parts of
Faces Using a Consensus of Exemplars,"
Peter N. Belhumeur, David W. Jacobs, David J. Kriegman, Neeraj Kumar,
Proceedings of the 24th IEEE
Conference on Computer Vision and Pattern Recognition (CVPR),
June 2011.
© Kriegman-Belhumeur Vision Technologies, LLC. 2011-2013