Metrics
After understanding the similarity score, the next step is to determine the acceptable similarity score threshold. This threshold is commonly known as the threshold value. If the face recognition similarity score is greater than the threshold, the verification will be accepted by the system. However, if the similarity score is less than the threshold, the verification will be rejected. To determine the threshold value, we need to determine the False Acceptance Rate (FAR) and False Rejection Rate (FRR) metrics.
To obtain the FAR and FRR values, we need to map the verification results based on the Confusion Matrix. The Confusion Matrix will help to assess the performance of the verification process, by mapping the predicted results to the actual situation. Here are the four determined conditions.
True Positive: This quadrant contains the number of times the model correctly predicted the positive class
False Positive: This quadrant contains the number of times the model incorrectly predicted the positive class
False Negative: This quadrant contains the number of times the model correctly predicted the negative class
True Negative: This quadrant contains the number of times the model incorrectly predicted the negative class
FAR and FRR
The performance of a face recognition system is generally determined by two types of error values: False Acceptance Rate (FAR) and False Rejection Rate (FRR).
False Acceptance Rate (FAR): the percentage of identification occurrences where an unauthorized person (a fake/different subject from the photo) is accepted by the system.
False Rejection Rate (FRR): the percentage of identification occurrences where an authorized person (the real subject/photo) is not accepted by the system.
If the determination of FAR and FRR values is not accurate, it will affect the level of security and user comfort. A higher FRR value will increase system security, but make it difficult for identified users to be accepted. On the other hand, if FAR is higher, the user identification process will be easier to accept, but it will make the system's security low. To avoid this, the lowest optimum point between FRR and FRR is sought, which is then called the Equal Error Rate (ERR) value. The ERR value is commonly used as the threshold value.
Threshold Value
The threshold value is determined by calculating FAR and FRR, so that the ERR value can be obtained and then set as the threshold. ERR can be used as the main reference as the threshold value because the FAR and FRR values are balanced. It is not a problem if the threshold value shifts from ERR, as long as the trade-off between FAR and FRR is considered. Do not choose a very small FAR value for the sake of tighter security, but make it easier for consumers to be rejected by the system and reduce ease of use.
Last updated
