Films Face Databases


This page describes and contains links to data sets we use to develop and test novel face recognition algorithms in the context of video retrieval and movie cast clustering.


This is a list of publications in which the datasets were used:

O. Arandjelović and R. Cipolla. Automatic cast listing in feature-length films with anisotropic manifold space. In Proc. IEEE Conference on Computer Vision Pattern Recognition, 2:1513-1520, June 2006. [LINK]
O. Arandjelović and A. Zisserman. On film character retrieval in feature-length films. In Interactive Video: Algorithms and Technologies, Springer-Verlag, ISBN 978-3-540-33214-5, 2007. [LINK]
O. Arandjelović and A. Zisserman. Automatic face recognition for film character retrieval in feature length films. In Proc. IEEE Conference on Computer Vision Pattern Recognition, 1:581-588, June 2005. [LINK]


In films "Pretty Woman" and "Groundhog Day" faces were automatically detected in every 10th frame using a local implementation of the Schneiderman-Kanade face detector and rescaled to the uniform scale of 121 by 121 pixels. In the "Open Government" episode of the "Yes, Minister" situation comedy, face detection was performed on every 5th frame, using the freely available implementation of the Viola-Jones cascaded detector, part of the Intel's OpenCV library. This data set contains a wider variation in pose than the former two. All three sets contain faces with a wide range of scale, expression and illumination changes, with frequent partial occlusions, see Fig. 1 and 2. Face images were converted to greyscale by averaging the R, G and B colour channels. A summary is shown in Fig. 3.

Schneiderman-Kanade face
detector (original paper):
Viola-Jones face detector
(original paper):
Intel's OpenCV library: [LINK]

Fig. 1. Principal recognition difficulties for the "Yes, Minister" data set.

(a) Groundhog Day(b) Pretty Woman
Fig. 2. Examples of Schneiderman-Kanade detected faces.

Data setDetectionsGround truthFace sizeRate
Groundhog Day5543No 121x121every 10 frames
Pretty Woman5582Yes 121x121every 10 frames
Yes, Minister7965No originalevery 5 frames
Fig. 3 Data sets summary.

Obtaining the datasets

Download "Pretty Woman" data(53.4 MB)[LINK]
Download "Groundhog Day" data(55.4 MB)[LINK]
Download "Yes, Minister"
("Open Government" episode) data
(102 MB)[LINK]