In order to the performances of two or more systems, measures like FAR, FRR, CMS are not sufficient since they are not so much correlated to the real operation context.
We should also consider other metrics, such as:
- Number of used dataset, in order to evaluate the generalizability of the system;
- Image size: larger and higher quality image provide more information but also more noise;
- Probe and Gallery size: the size must be reasonably big in order to guarantee generalizability;
- Number and level of tolerated variations: this also guarantees robustness and generalizability, since in most of the cases for public systems, we cannot take for granted that a person will always appear the same.
FERET program
The FERET (Facial Evaluation Recognition Technology) was a program invented in 1993 that aimed to create a large and automatic face-recognition system. Before that date, a lot of facial recognition systems claimed a very high accuracy on a small sized database, but few of these systems used a common database.
The first evaluation on the FERET database took place in the August of 1994, the second in March 1995, the third on September 1996.
FRVT - Face Recognition Vendor Test
Note
Missing