How do you evaluate Facial recognition/verification technology?

It all depends on the use case for which facial recognition/verification technology is used. For example, a facial recognition algorithm for immigration is a 1:1 algorithm, while a facial recognition algorithm used in security cameras for general surveillance is a 1:n algorithm. Additionally, based on the use case, the threshold settings are configured. For example, during an access control operation, you may want the false match rate to be low, even when the false non-match rate is high, as you do not wish anyone who is not authorised to enter the premises. You would also want to check for liveness detection to ensure that the right person is entering the premises. On the contrary, for counter-terrorist operations, you may not want to miss a suspect, and the faces are scanned in the wild, i.e. in general public places on an edge device. In brief, different use cases require different sets of evaluation parameters.

?In brief different use-cases require a different set of evaluation parameters.

?E-KYC

a.????Good User experience: The ability to verify on a mobile device under different lighting conditions would significantly improve the user experience and ease of use.

b.????Accuracy: An excellent facial recognition algorithm that performs well on NIST for 1:1 category, the algorithm needs to be evaluated for different ages, as for most banks, the photographs on file will be more than ten years old.

c.?????Racial Bias: It is good to evaluate the algorithms under different lighting conditions with your data set after checking how accurate the algorithm is on NIST for racial bias.

d.????Liveness Detection: It is essential to have a good liveness detection algorithm for account opening use cases.

e.????Data Privacy: Always reading the fine print on how the data will be stored or used in training the algorithm would reduce inconvenience for all parties.

Access Control:

a.????Accuracy: An excellent facial recognition algorithm that performs well on NIST for 1:1 category, the algorithm needs to be evaluated for different ages as, for most organisations, the photographs on file will be more than ten years old. Another important criterion is the makeup, especially for women, as a higher threshold setting might result in denying access.

b.????Racial Bias: It is good to evaluate the algorithms under different lighting conditions with your data set after checking how accurate the algorithm performs on NIST for racial bias.

c.?????Liveness Detection: It is essential to have a good liveness detection algorithm for account opening use cases.

d.????Data Privacy: Always reading the fine print on how the data will be stored or used in training the algorithm would reduce inconvenience for all parties.

Security

a.????Accuracy: An excellent facial recognition algorithm that performs well on NIST for 1:n category for wild images.

b.??????Racial Bias: It is good to evaluate the algorithms under different lighting conditions with your data set after checking how accurate the algorithm performs on NIST for racial bias.

c.????Speed: The speed of recognition is an essential criterion as people are moving, and you might miss the suspect if the algorithm is too slow in recognising.

d.????Data Privacy: Always reading the fine print on how the data will be stored or used in training the algorithm would reduce inconvenience for all parties.

e.??????Camera type: The camera should preferably use a normal lens for surveillance and not a wide-angle lens; it is preferred to have a wdr enabled camera as they automatically adjust to different lighting settings.

f.??????Mounting location: This is another critical aspect, as it is important to mount it such that the pitch, yaw, and roll are within the limits of the facial recognition algorithm

g.?????Search capability: For surveillance operations, the cameras need to be integrated as this would allow post-analysis in case the need arises

h.??????Installation of the cameras: Installing the cameras at the chokepoints ensures everyone is scanned.

i.??Occlusions; Ability to recognise a person of interest with occlusions is critical.

j.????Support: It is an important aspect that every facial recognition company needs to recognise its importance.