EYE COORDINATES LOCATED USING ROBUST CORRELATION, XM2VTSDB DATABASE
-------------------------------------------------------------------

 File: README_RC
 Author: Kenneth Jonsson, Univ of Surrey
 Date: Tue Feb  2 10:02:05 GMT 1999
 Updated: Tue Feb  2 10:02:05 GMT 1999

DATA FILES:        NOTES:
-----------        ------

 xm_c1_ce_rc.dat - Client evaluation set
 xm_c1_ct_rc.dat - Client test set
 xm_c1_ie_rc.dat - Impostor evaluation set
 xm_c1_it_rc.dat - Impostor test set
 xm_c1_tr_rc.dat - Training set

 See README_MANUAL for file name conventions.

FORMAT OF XM2VTSDB DATA FILES:
------------------------------

 xxx_x_x xxx xxx xxx xxx
  |  | |  |   |   |   |_ Right eye, y coord
  |  | |  |   |   |_____ Right eye, x coord
  |  | |  |   |_________ Left eye, y coord
  |  | |  |_____________ Left eye, x coord
  |  | |________________ Shot
  |  |__________________ Session
  |_____________________ Subject ID

SOME NOTES ON THE LOCALISATION AND ACCURACY OF THE COORDINATES:
---------------------------------------------------------------

- The prediction was done by locating the coordinates in one 
  image (371_2_1, the root frame of the registration tree) 
  manually and then predicting the coordinates in the remaining 
  ones.
- Many of the impostor eye-coordinates are weird (maybe even
  outside the image) which is a consequence of the fact that
  you're trying to register images of different people. The
  robust correlation was not designed to handle this.
- For impostors the transformations obtained using the robust
  correlation depends on the claimed identity. Therefore, in 
  order to establish a unique set of coordinates, the identity 
  of the root frame was selected. 
- The coordinates are given in the original image resolution. 
- The eye coordinates were initially located at a lower
  resolution (the robust correlation is a multi-resolution
  approach) and when the coordinates are extrapolated to the
  original resolution there is some loss of accuracy. This
  is due to the fact that the pixels are shifted slightly
  when the images are smoothed and subsampled.
- For more information about the method used to located the eye
  coordinates refer to 
 
   K. Jonsson, J. Matas and J. Kittler
   "Learning Salient Features for Real-Time Face Verification"
   To be presented at Audio and Video Based Biometric Person 
   Authentication (AVBPA'99), March 1999.