Description of the face database

The dataset consists of 1521 gray level images with a resolution of 384×286 pixel. Each one shows the frontal view of a face of one out of 23 different test persons. For comparison reasons the set also contains manually set eye postions. The images are labeled “BioID_xxxx.pgm” where the characters xxxx are replaced by the index of the current image (with leading zeros). Similar to this, the files “BioID_xxxx.eye” contain the eye positions for the corresponding images.

Image file format

The images are stored in single files using the portable gray map (pgm) data format. A pgm file contains a data header followed by the image data. In our case the header consists of four lines of text. In detail:

• the first line describes the format of the image data (ASCII/binary). In our files the text “P5” indicates that the data is written in binary form

• the second line contains the image width written in text form

• the third line keeps the image height also in text form

• the fourth line contains the maximum allowed gray value (255 in our images)

The header is followed by a data block containing the image data. The data is stored line per line from top to bottom using one byte per pixel.

Eye position file format

The eye position files are text files containing a single comment line followed by the x and the y coordinate of the left eye and the x and the y coordinate of the right eye separated by spaces. Note that we refer to the left eye as the person’s left eye. Therefore, when captured by a camera, the position of the left eye is on the image’s right and vice versa.

Evaluation of face detection algorithms

To make it possible to compare the quality of different face detection algorithms on the testset we propose the following distance-based quality measure:

• Estimate the eye positions with your algorithm and calculate the absolute pixel distance from the manually set positions so that you receive two distance values.

• Choose the larger value and divide by the absolute pixel distance of the two manually set eye positions so that you become independent from the face’s size in the image. We call this value relative eye distance.

• When calculating this distance for each image you can choose the distribution function of the relative distances to compare some results with others.