Enrollment
In order to perform face-recognition (either verification or identification) you first need to perform an Enrollment. That is, a description for a subject from their detected faces in the input image(s) or video frames must be generated. In face-recognition this description is more commonly known as a biometric template. Also, the process of turning a detected face(s) into a description is often referred to as feature-extraction with the data held in the description referred to as the features.
In the early days of face-recognition these features were handcrafted and measured things like the distance between the eyes or the length of the nose. However, with modern algorithms these features are implicitly generated by powerful machine learning algorithms.
The diagram below shows the process for generating a description for a subject from a set of images or a video.
Matching
In order to compare the identities of faces in two different images or video, you need to first compute a description for each of them. Verification is the process of matching two descriptions against each other to determine whether they belong to the same identity.
In order to compare two descriptions a comparer algorithm is used which is capable of comparing the information held in the descriptions. The job of this comparer algorithm is to provide a match score which determines how similar the two descriptions are. The higher the match score then the more probable the identity of the two faces will be the same.
The comparer used by Digital Barriers can output scores in the range of -1.0 to 1.0. A score of 1.0 represents a perfect match and is only achieved when the two descriptions are numerically equivalent. However, in normal operation, the scores lie in the range of -0.1 and 0.85.
Note, the match score is not a percentage value and does not map onto a 0 to 1 scale. Also, there are different comparer algorithms and others can have a completely different output range.
Thresholds
For most face-recognition applications, a threshold value must be set. The match score is then compared against this threshold and if it exceeds this value, the faces represented by the descriptions are considered the same identity. If the match score is less than the threshold, the faces represented by the descriptions are not considered to be of the same identity. The setting of this threshold value depends on the security requirements and desired error rates for your application.
The table below gives suggested threshold levels for different error rates. By default, for SmartVis Face Applications (i.e Mobile, Server and Identifier) a threshold of 0.6 is used. You can see that this in fact a very high threshold value. This helps limit the False Positives.
The table below gives the expected False Rejection Rates at the different False Acceptance Rates, given by the thresholds in the table above.
More information on these Error Rates and threshold levels can be found in our Face Recognition developer documentation.