For the manual part to the ICBA and ICPR Face Autentication Competitions a set of geometrically normalised images have been supplied.
The geometric normalisation used consists of two steps.
- Smoothing
- Rotating and scaling
The smoothing is performed using a Gaussian Filter of 5x5 which is given by the convolution of the filter:
The rotation and scaling outputs an image of size 55 rows x 51 cols. The left-eye is mapped onto the pixel position (19, 38) and the right-eye is mapped onto the pixel position (19, 12). The diagram below shows a representation of this.
I have made the source code that performs the geometric normalisation available. However, it relies on the RAVL image libraries..
I have also compiled a linux binary that should work on most linux i386 architectures.