[ Preprocessing for View-Based Systems ]
Computer Vision      
  Preprocessing for View-Based Systems.  

Most face recognition systems that are currently available are restricted to frontal views of faces, and their performance drops significantly if the faces are slightly rotated. We have presented a method to generate frontal views from non-frontal images of faces, based on 3D face reconstruction from single images. This can be included as a preprocessing step in face recognition frameworks to combine the efficiency of commercial, view-based systems with the flexibility of our 3D model-based approach.

  Image data generated by our algorithm have been part of the Face Recognition Vendor Test FRVT2002 by the National Institute of Standards and technology (NIST) and DARPA. The test set included frontal views and left, right, up and down views of 87 individuals. 10 commercial face recognition systems were given the set of frontal views as gallery of known faces. Tested on the original, rotated faces, the rate of correct verification of the best systems was between 40% (left views) and 55% (right and down views) at false accept rate of 1%. Tested on the images preprocessed by our system, verification rate increased to 79% - 85% for the best systems. Nine out of ten systems' performance increased, often dramatically, in all testing conditions.  
Blanz, V. and Vetter, T., Generating Frontal Views from Single, Non-Frontal Images,
In: Face Recognition Vendor Test 2002, Technical Appendices,
Nat. Inst. of Standards and Technology (NIST),
NISTIR 6965, Appendix O, 2003. pdf

© Sascha Nesch 2009