Robert's blog
Robert Važan

Speed of fingerprint recognition doesn't matter

A few back-of-the-envelope calculations show that overoptimizing fingerprint matchers is useless and brings no benefits to end users.

Matching speed is obviously irrelevant for small databases, but I am going to show you that it doesn't matter on large databases either.

Rented cloud node running SourceAFIS will do 10K fingerprints in 2.5 seconds, which is an acceptable delay for the large-scale applications. That means we need N/10K nodes to search DB of N fingerprints, e.g. 100 nodes to search 1M fingerprints.

A small, single-core cloud server kept alive only 40 hours/week costs about $1 per month when using per-hour billing. The cost of searching database of N fingerprints is N/10K in dollars, e.g. $100/mo to be able to search 1M fingerprints in 2.5 seconds.

Typically the number of searches (identifications) per month is 1/10th of database size, e.g. 100K searches monthly in 1M database. That translates to $0.001 (tenth of a cent) per identification request ($100 divided by 100K searches).

With 40 hours/week, we have about 200K time slots on every cloud node, every month, each lasting 2.5s. That means we have to start shortening time slots with databases of more than 2M fingerprints, which have over 200K identifications/month.

That means we get price of $0.001 per identification for searches of up to 2M fingerprints. Then the cost scales up linearly with database size, e.g. $0.05/identification on database of 100M fingerprints.

Even that price can be reduced with a few tweaks in the algorithm that increase its speed 10x at the cost of a minor increase of FRR.

Since typical biometric vendor will charge you no less than $0.01 in license fees for every enrolled user in the database, it is clear that hardware cost is negligible even with 10K fp/s algorithm. Unless you are searching fingerprints of the whole population of India, of course.

The reason for driving up matching speeds in biometric industry was to keep the number of servers low, because most deployments in the past have been based on obsolete concept of in-house server room, which of course couldn't be as cheap as modern cloud services.

That's why I see no reason to drive up matching speed in SourceAFIS. It's much wiser to invest in ease of use and accuracy. But then, accuracy is a dubious metric too. It really comes down to what software is easiest to use within given price category.