Labeled Faces in the Wild 人脸识别数据集

阅读量：2117 次

发布时间：2019-04-30

本文共 15813 字，大约阅读时间需要 52 分钟。

Labeled Faces in the Wild Home

New (draft) survey paper:

Erik Learned-Miller, Gary Huang, Aruni RoyChowdhury, Haoxiang Li, Gang Hua

The camera-ready has not yet been submitted. If you see typos or errors, please let us know and we will try to correct them.

New results page:

We have recently updated and changed the format and content of our . Please refer to the for details of the changes.

Welcome to Labeled Faces in the Wild, a database of face photographs designed for studying the problem of unconstrained face recognition. The data set contains more than 13,000 images of faces collected from the web. Each face has been labeled with the name of the person pictured. 1680 of the people pictured have two or more distinct photos in the data set. The only constraint on these faces is that they were detected by the Viola-Jones face detector. More details can be found in the technical report below.

There are now four different sets of LFW images including the original and three different types of "aligned" images. The aligned images include "funneled images" (ICCV 2007), LFW-a, which uses an unpublished method of alignment, and "deep funneled" images (NIPS 2012). Among these, LFW-a and the deep funneled images produce superior results for most face verification algorithms over the original images and over the funneled images (ICCV 2007).

[new] .

LFW .

LFW (see ).

, our new database for face detection research.

workshop at the , run by Erik Learned-Miller, Andras Ferencz, and Frederic Jurie.

last updated: 2015/09/22 15:09 EDT

If you wish to receive announcements regarding any changes made to the LFW database, please send email to (subject and body are ignored).

Alphabetically by first name:

Alphabetically by first name, only people with more than one image:

Alphabetically by last name:

By number of images per person:

(173MB, md5sum a17d05bd522c52d84eca14327a23d494)

[new]
(111MB, md5sum 68331da3eb755a505a502b5aacb3c201)
- see .

(233MB, md5sum 1b42dfed7d15c9b2dd63d5e5840c86ad)
- see .

(LFW-a - Taigman, Wolf, Hassner)
See also LFW3D (frontalized LFW images) under below.

Superpixel segmentations:
- (328MB, md5sum eb6543ba9bbef54f8ba481c895d3526f)
- (129MB, md5sum 5a166aa967e260aa70d55b5785aa7a61)
- (328MB, md5sum f1ede21969d2ad8262a16a26d6212177)

To download LFW attribute values (Attribute and Simile Classifiers for Face Verification, Kumar et al.), see the .

(14MB)

(individual person with most images) (6.9MB)

- information on file formats and directory structure

View 1:

For development purposes, we recommend using the below training/testing split, which was generated randomly and independently of the splits for 10-fold cross validation, to avoid unfairly overfitting to the sets above during development. For instance, these sets may be viewed as a model selection set and a validation set. See the tech report below for more details.

Explore the sets:

Download the sets: , , ,

View 2:

As a benchmark for comparison, we suggest reporting performance as 10-fold cross validation using splits we have randomly generated.

Explore the sets:

Download the sets: ,

For information on the file formats, please refer to the README above.

For details on how the sets were created, please refer to the tech report below.

Accuracy and ROC curves for various methods available on .

13233 images

5749 people

1680 people with two or more images

The following is a list of known errors in LFW. Due to the small number of such errors, the database will be left as is (without corrections) to avoid confusion.

It is important that users of the database provide their algorithms with the database

as is, i.e. without correcting the errors below, since previous results published for the database did not have the advantage of correcting for these errors.

Currently, there are five incorrectly labeled matched pairs in View 2. While we do not believe this should have a significant effect on accuracy, we do encourage researchers to be aware of these errors when producing any visualizations (e.g. matched pairs most confidently predicted as mismatched, as the matched pair may actually be mismatched).

The current known errors in View 2 are:

Fold 1: Janica_Kostelic_0001, Janica_Kostelic_0002

Fold 1: Nora_Bendijo_0001, Nora_Bendijo_0002

Fold 5: Jim_OBrien_0001, Jim_OBrien_0002

Fold 5: Jim_OBrien_0001, Jim_OBrien_0003

Fold 5: Elisabeth_Schumacher_0001, Elisabeth_Schumacher_0002

More detail about all the errors is given below.

Note: unless stated otherwise below, any error in a matched pair will mean that the label ("matched") is wrong. Any error in a mismatched pair, even with the person having the wrong identity, will generally be correct (the label of "mismatched" will still be correct).

_0004 is incorrect (it is an image of Abdullah Gul):

This image appears only in one matched pair in the training set of View 1:

, 2 , 4

_0001 is incorrect (it is an image of Janica Kostelic):

This image does not appear in a matched or mismatched pair, in either view.

_0001 is incorrect (it is an image of Anja Paerson):

This image appears in one matched pair in the test set of View 1, and the same matched pair and one mismatched pair (with Don_Carcieri_0001) in fold 1 of View 2:

, 1 , 2

_0001 is incorrect (it is a duplicate image of Ricky_Ray_0001):

This image appears in two mismatched pairs in the training set of View 1, and one mismatched pair in fold 2 of View 2. (None of the mismatched pairs are with Ricky_Ray.)

_0001 is incorrect (it is a duplicate image of Raul_Ibanez_0001):

This image appears in one mismatched pair in the test set of View 1, and one mismatched pair in fold 5 of View 2. (None of the mismatched pairs are with Raul_Ibanez.)

_0001 is incorrect (it is a duplicate image of Eva_Amurri_0001):

This image appears in one mismatched pair in the test set of View 1 (the mismatched pair is not with Eva_Amurri).

_0008 is incorrect (it is an image of Rubens Barrichello):

This image does not appear in a matched or mismatched pair, in either view.

_0012 is incorrect (it is an image of Hamad Bin Isa al-Khalifa):

This image does not appear in a matched or mismatched pair, in either view.

contains two distinct persons. Specifically, Jim_OBrien_0001 is a different person from Jim_OBrien_0002, Jim_OBrien_0003.

This leads to an error in two matched pairs (0001 with 0002; 0001 with 0003), present in both the training set of View 1 and fold 5 of View 2:

, 1 , 2

, 1 , 3

_0001 is an incorrect spelling of .

This image appears in a mismatched pair in fold 3 of View 2 (not with Jon_Gruden).

contains two distinct persons, where the correct spelling of Elisabeth_Schumacher_0001 is actually Elizabeth Schumacher. This leads to a incorrect matched pair in both the test set of View 1 and fold 5 of View 2.

, 1 , 2

_0001 is actually an image of .

This image appears in a mismatched pair in the training set of View 1 and in fold 1 of View 2 (neither with Andrew_Caldecott).

_0002 is actually an image of , and_0002 is actually an image of

, 1
correct label , 2
is actually Flor Montulo

, 1
correct label , 2
is actually Nora Bendijo

Nora_Bendijo_0002 appears in an incorrect matched pair in fold 1 of View 2.
Flor_Montulo_0002 appears in an incorrect matched pair in the test set of View 1, and in a mismatched pair in fold 1 of View 2, but not with Nora_Bendijo.

and are the same person. The two names are never together in a mismatched pair.

and are the same person. The two names are never together in a mismatched pair.

and are the same person.

These images never appear in a mismatched pair together.

_0001 is actually an image of .

This image never appears in a mismatched pair with an image of Takahiro_Mori.

Please cite as:

, Manu Ramesh, , and .

Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments.

University of Massachusetts, Amherst, Technical Report 07-49, October, 2007.

BibTeX entry:

@TechReport{LFWTech,  author =       {Gary B. Huang and Manu Ramesh and Tamara Berg and                   Erik Learned-Miller},  title =        {Labeled Faces in the Wild: A Database for Studying                   Face Recognition in Unconstrained Environments},  institution =  {University of Massachusetts, Amherst},  year =         2007,  number =       {07-49},  month =        {October}}

and .

Labeled Faces in the Wild: Updates and New Reporting Procedures.

University of Massachusetts, Amherst, Technical Report UM-CS-2014-003, May, 2014.

@TechReport{LFWTechUpdate,  author =       {Gary B. Huang Erik Learned-Miller},  title =        {Labeled Faces in the Wild: Updates and New Reporting                   Procedures},  institution =  {University of Massachusetts, Amherst},  year =         2014,  number =       {UM-CS-2014-003},  month =        {May}}

If you use the LFW images aligned by funneling, please cite:

Gary B. Huang, , and .

International Conference on Computer Vision (ICCV), 2007.

@InProceedings{Huang2007a,  author =    {Gary B. Huang and Vidit Jain and Erik Learned-Miller},  title =     {Unsupervised Joint Alignment of Complex Images},  booktitle = {ICCV},  year =      {2007}}

If you use the LFW imaged aligned by deep funneling, please cite:

Gary B. Huang, , , and .

Advances in Neural Information Processing Systems (NIPS), 2012.

@InProceedings{Huang2012a,  author =    {Gary B. Huang and Marwan Mattar and Honglak Lee and                Erik Learned-Miller},  title =     {Learning to Align from Scratch},  booktitle = {NIPS},  year =      {2012}}

Collected resources related to LFW:

Note: We have not verified the accuracy or reliability of the code and data at the following links; we merely provide them as a convenience. Please use your own judgment about the accuracy of the resources below.

"Getting the known gender based on name of each image in the Labeled Faces in the Wild dataset. This is a python script that calls the genderize.io API with the first name of the person in the image."

"While there are many open source implementations of CNN, none of large scale face dataset is publicly available. The current situation in the field of face recognition is that data is more important than algorithm. To solve this problem, we propose a semi-automatical way to collect face images from Internet and build a large scale dataset containing 10,575 subjects and 494,414 images, called CASIA-WebFace. To the best of our knowledge, the size of this dataset rank second in the literature, only smaller than the private dataset of Facebook (SCF). We encourage those data-consuming methods training on this dataset and reporting performance on LFW. "

- collection of frontalized LFW images and Matlab code for frontalization

"Frontalization is the process of synthesizing frontal facing views of faces appearing in single unconstrained photos. Recent reports have suggested that this process may substantially boost the performance of face recognition systems... we explore the simpler approach of using a single, unmodified, 3D surface as an approximation to the shape of all input faces. We show that this leads to a straightforward, efficient and easy to implement method for frontalization. More importantly, it produces aesthetic new frontal views and is surprisingly effective when used for face recognition and gender estimation."

Questions and comments can be sent to:

The building of the LFW database was supported by NSF CAREER Award number 0546666.

2015/09/22

Added Ding et al.

, Ding and Tao

, and Xu et al.

to .

2015/07/31

Updated . Added Baidu, AuthenMetric commercial system results to .

2015/06/03

Updated .

2015/05/19

Updated .

2015/04/13

Added Sun et al.

, Li and Hua

, updated betaface.com commercial system

, on .

2015/01/19

Updated Face++ commercial system result

, added betaface.com commercial system result

2014/12/09

Added Ouamane et al.

and Sun et al.

to .

2014/12/08

Added new .

Added Hassner et al.

and TCIT

to .

2014/12/03

Added Hu et al.

to .

2014/12/01

Added Arashloo and Kittler

to .

2014/09/19

Added Li et al.

to .

2014/06/20

Added Sun et al.

to .

2014/06/16

Added Lu and Tang

to .

2014/06/12

Added Sun et al.

to .

2014/05/28

Added AUC for two unsupervised methods - LHS and MRF-MLBP - to .

2014/05/23

Added Zhu et al.

to .

2014/05/22

Updated Kumar et al.

with results from journal paper, and added Berg and Belhumeur

, on .

2014/05/21

Added Sun et al.

and Hu et al.

to .

2014/05/19

Added Taigman et al.

to .

2014/05/09

Re-organized to more accurately and fairly compare algorithms under a variety of protocols. See the for details of the changes.

2014/03/17

Added Face++ commercial system result

2014/02/10

Added Aurora Computer Services commercial system result

2014/01/09

Added VisionLabs commercial system result

2013/12/08

Added John_Gruden_0001 labeling error to .

2013/10/18

Added Cao et al.

, Cao et al.

, and Barkan et al.

to .

2013/10/08

Added Lei et al.

to .

2013/08/23

Added new deep funneled images and corresponding deep funneled superpixels for download. (CJG)

2013/08/12

Temporarily disabled deep funneled data set download. (CJG)

2013/08/09

Removed superpixel images from database gallery. Added deep funnel aligned images to database gallery. (CJG)

2013/08/08

Added deep funneled aligned image set to main page. (CJG)

2013/07/28

Added Simonyan et al.

to .

2013/07/13

Added Yi et al.

to .

2013/07/13

Added Arashloo and Kittler

to .

2013/06/12

Added Sharma et al.

to .

2013/05/12

Added Zhen et al.

to .

2013/05/02

Added Chen et al.

and Chen et al.

to .

2013/04/17

Added Anja_Paerson_0001 labeling error to .

2013/03/21

Added Li et al.

to .

2012/08/17

Added Hussain et al.

to .

2012/07/18

Added Berg and Belhumeur

to .

2012/03/03

Added Huang et al.

to .

2011/12/14

Added Ying and Li

to .

2011/09/07

Added link to download computed attribute values for all LFW images produced by Kumar et al., on the .

2011/08/28

Added Seo and Milanfar

to .

2011/08/08

Added images of incorrectly labeled faces, in .

2011/08/08

Added Taigman and Wolf

to .

2011/08/01

Added Jim_OBrien_0001 labeling error to .

2011/07/18

Updated the , adding notes on the use of external training data, arranging the image-restricted method results to roughly reflect the amount of external training data used, and added specific notes on the type of external training data used for each algorithm.

2011/07/12

Added Mahmoud_Abbas_0012 labeling error to .

2011/06/28

Added Yin et al.

to .

2011/04/28

Added superpixel segmentation files to .

2011/04/04

Added Li et al.

to .

2011/01/29

Added Pinto and Cox