Face Localization - Hierarchical Neural Networks for Image Interpretation - page 202

Fig. 10.2. Some face images from the BioID dataset. Since the examples vary considerably,
the dataset can be considered challenging.
scanning through parameter spaces is needed, and multiple detection candidates co-
operate and compete with each other to produce a coherent image interpretation.
10.2 Face Database and Preprocessing
To validate the performance of the proposed approach for learning face localization,
the BioID database [111] is used. The database can be downloaded free of charge
from http://www.bioid.com/downloads/facedb/facedatabase.html . It consists of 1,521
images that show 23 individuals in front of various complex office backgrounds
with uncontrolled lighting. The persons differ in gender, age, and skin color. Some
of them wear glasses and some have beards. Since the face size, position, and view,
as well as the facial expression vary considerably, the dataset can be considered
challenging.
Such real-world conditions are the ones that show the limits of current local-
ization techniques. For instance, while the hybrid localization system, described
in [111], correctly localizes 98.4% of the XM2VTS database [157] which has been
produced under controlled conditions, the same system localizes only 91.8% of the
BioID faces. Figure 10.2 shows some example images from the BioID dataset.
The gray-scale BioID images have a size of 384 × 288 pixels. To reduce border
effects, the contrast is lowered towards the sides of the image. To limit the amount
of data, the image is subsampled to 48 × 36, 24 × 18, and 12 × 9 pixels, as shown in
Figure 10.3(b). In addition to the images, manually labeled eye positions C l ,C r ∈
R 2 are available. They are in general quite reliable but not always as accurate as one
could hope.
Figure 10.3(a) shows the marked eye positions for a sample image. A multi-
resolutional Gaussian blob is produced for each eye in a set of images that have the
above resolutions. The blobs are shown in Figure 10.3(b). Their standard deviation
σ is proportional to the distance of the eyes k C l − C r k . Note that with increasing
resolution, the area of the blob increases, with respect to the original image. Thus,