Main Page
From faces
Contents |
Introduction
The primary reason for creating this wiki was a wish to summarize the progress as well as current state of facial recognition systems in the world. All the information on this site is my personal experience and opinion. If You find some inconsistencies or global errors in my summary please contact me by justas[eta]atrodo[dot]com with Your evidentiary material. Basic knowledge of mathematics, linear algebra, applied statistics, image processing or similar (like broken English) is needed to understand my point of view.
For example, every single "survey" on face recognition starts from eigenfaces as a breakthrough in appearance based methods. Well, that was true in 1991 but now i can cite the astonishment of my college: "Are you kidding? It just doesn't work!" Nonetheless, 90% of all publications still do a research in PCA (Principle Component Analysis) modifications that improve the results of original eigenfaces method. Additionally, every new method (not only the PCA modifications) is compared to eigenfaces and show smaller or larger improvements in quality of recognition. Even FRGC (Face Recognition Grand Challenge) uses eigenfaces as a baseline algorithm while average face recognition algorithm is five to ten times more precise. Why do they do that? It is classic. It is simple. It is fast to implement. And, in very rare occasions it even works. On the other hand, would number of publications be less quantitative and more qualitive if some new baseline algorithm was developed for comparisons?.. No.
That's why this summary will be oriented towards commercial systems and independent testing. Of course, basic concepts, methods and ideas will be reviewed, explained or even implemented.
Face Recognition
Let's start from the "simplest" face recognition. Actually, it is simple to define the problem but not to solve it:
- Given two images with faces, tell whether faces belong to the same person.
It's another story what recognition means but in most general case the face recognition algorithm should be able to compare two images with faces and tell whether these faces are same or different. Most of the algorithms do not make direct comparison of images. Firstly, they extract some meaningful and discriminating features of faces from them. Secondly, they compare previously extracted features. Additionally, most of the algorithms do not return a discrete answer - true or false - when compare two faces, mostly, they return a similarity score between the faces:
(image1) -> (face detection) -> (feature extraction) -. 1.0 (same) \ . > (compare) -> . / . (image2) -> (face detection) -> (feature extraction) -' 0.0 (different)
The most discriminating and invariant features are extracted for fast comparison, small memory footprint and best quality of recognition (smallest number of errors).
TODO: face detection, normalization, recognition.
TODO: 2D, 3D, fusion.
TODO: low resolution, high resolution, multiple stills.
Vendors
All information was gathered from the web. Contact me by justas[eta]atrodo[dot]com if You would like to update the information.
| company | minimum IOD1 | face localization2 | face tracking3 | single image processing4 | template size5 | matching speed6 | pose variation7 |
|---|---|---|---|---|---|---|---|
| Cognitec Systems Gmbh (Core 2 Duo, 2.8 Ghz) | 32 | - | 40 | 100 | 800 (A13T7) 1800 (B3T7) | 230 000 | - |
| L-1 Identity Solutions (Pentium 4, 2.0 Ghz) | 60 | 300 | - | 800 | 648 (VFA) 5325 (LFA) 6861 (STA) | 1 000 000 (VFA) 160 000 (VFA+LFA+STA) | up to 15 (not significant) 15 to 35 (slight) 35 to 90 (significant) |
| Neurotechnology (Pentium 4, 3.0 Ghz) | 50 | 70 (MegaMatcher) 70 (VeriLook) 45 (FaceCell) | - | 200 (MegaMatcher) 200 (VeriLook) 60 (FaceCell) | 2356 | 500 000 (MegaMatcher) 200 000 (VeriLook) 300 000 (FaceCell) | up to 5 (not significant) 5 to 10 (slight) |
| Acsys Biometrics Corp. | 40 (Mugshot) 15 (All Aspect) | - | 33 | 1000 | 4096 (Mugshot) 8192 (All Aspect) | 100 000 (Mugshot) 25 000 (All Aspect) | up to 15 (Mugshot) up to 90 (All Aspect) |
1 minimum inter-occular distance (in pixels)
2 face localization time (in ms)
3 face tracking time for one frame (in ms)
4 single image processing (extraction of one face template from one image) time (in ms)
5 template size (in bytes)
6 matching speed (in face templates per second)
7 matching ability degradation when face pose is varying (in degrees)
History
While reading the FRVT reports and doing research in facial recognition history i thought it would be interesting to show the path of evolution of some companies in the following table. It was inspired by FERET Transition image. Some information on technologies used in their recognition algorithms still did not fade from the internet and should be added into this table together with more detailed list of companies/technologies acquired by them.
| Google Inc. | L-1 Identity Solutions | Cognitec | |||||||
|---|---|---|---|---|---|---|---|---|---|
| 1994 | USC, Laurenz Wiskott | 1994 | USC, Christoph von der Malsburg | MIT, Alex Pentland | Rockefeller University, Joseph Atick Visionics Technology | 1994 | 1994 | ||
| 1995 | 1995 | 1995 | Siemens Nixdorff | 1995 | |||||
| 1996 | 1996 | 1996 | 1996 | ||||||
| 1997 | Eyematic Interfaces | 1997 | Lau Technologies | 1997 | 1997 | ||||
| 1998 | 1998 | 1998 | 1998 | ||||||
| 1999 | 1999 | Provista | 1999 | 1999 | |||||
| 2000 | 2000 | ZN Vision Technologies AG | Viisage Technology | 2000 | 2000 | ||||
| 2001 | 2001 | Delean Vision, Bruno Delean | Visionics Corporation | 2001 | 2001 | ||||
| 2002 | 2002 | merged with Identix | 2002 | Cognitec Systems Gmbh | 2002 | ||||
| 2003 | Neven Vision | 2003 | 2003 | 2003 | |||||
| 2004 | 2004 | Viisage | Identix | 2004 | 2004 | ||||
| 2005 | 2005 | 2005 | 2005 | ||||||
| 2006 | Google Inc. | 2006 | L-1 Identity Solutions | 2006 | 2006 | ||||
| 2007 | 2007 | 2007 | 2007 | ||||||
| 2008 | 2008 | 2008 | 2008 | ||||||
Evaluation
TODO: identification, verification, watchlist
Databases
TODO: feret, yale a, yale b, pie, bioid, xm2vts, banca, at&t (orl).
