Caltech Vision Lab

Caltech 10, 000 Web Faces

[Description] [ Download] [ References ] [Resources]



The dataset contains images of people collected from the web by typing common given names into  Google Image Search. The coordinates of the eyes, the nose and the center of the mouth for each frontal face are provided in a ground truth file. This information can be used to align and crop the human faces or as a ground truth for a face detection algorithm. The dataset has 10,524 human faces of various resolutions and in different settings, e.g. portrait images, groups of people, etc. Profile faces or very low resolution faces are not labeled.


Before you download the data, please note: The pictures in the dataset were harvested from the web for the purpose of carrying out not-for-profit scientific experiments and are not Caltech property. Any use of the dataset, other than  'fair use', must be negotiated with the pictures' owners. Caltech is not responsible for the content nor the meaning of the images.

Web faces dataset : Caltech_WebFaces.tar [139Mbytes]

Face features ground truth : WebFaces_GroundThruth.txt ReadMe.txt    


Papers using the dataset:

Anelia Angelova, Yaser Abu-Mostafa, Pietro Perona, Pruning Training Sets for Learning of Object Categories , Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2005

See also the Data Pruning project webpage

Michael Fink, Pietro Perona, Mutual Boosting for Contextual Inference, Proc. Neural Information Processing Systems (NIPS), 2003


Here are a number of useful matlab scripts for this data:

A matlab script which displays the images is here.


The data contains a total of 10,524 faces in 7,092 images. The average image resolution is 304x312 pixels across the data. Here is a script which displays image resolution statistics.

The statistics of the resolution of faces present in the dataset are presented below. The matlab script used to generate them is here.

FacesPerImage EyeDist

The data has a number of duplicate images. These duplicate images were distributed among different people to provide ground truth and can be used to evaluate reliability and precision of the manually generated ground truth. Here is a list of the images which we believe to be duplicates and here is a script which identifies them.


The images and the ground truth were collected and organized by  Michael Fink while visiting the Caltech Vision Group . Rob Fergus provided a script for harvesting images from Google.


If you have any questions regarding the data please contact: 

Anelia Angelova

anelia [at] vision [.] caltech [.] edu

CS Department, California Institute of Technology

1200 E. California Blvd. MC 136-93, Pasadena, CA, 91125, USA

This site is maintained by Anelia Angelova. Last updated: February 14, 2007.