Ask Your Question

Facial Recognition Speed

asked 2015-08-19 10:01:00 -0500

DanG gravatar image

Hi guys,

I'm a bit of a noob with facial recognition, but have a question I was hoping someone could answer regarding speed.

If we take a database of 100,000 users, each with 1 photo (forward facing); how long will it take for the results to return?

I'm thinking of likeness based on 91-100% match and then 80-90%. This means there will be more than one result returned.

So, a user scans a photo into the database and gets returned, say, 5 images based on 91-100% match and then 10 images based on 81-90%

In terms of the server running these queries, the specs would be a quad core i7 with 16GB ram.

What I'm looking for are rough estimates i.e. will the response take seconds, minutes, hours etc.

I'm sure there's probably a few variables I've missed, but any help would be appreciated.


edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted

answered 2015-08-19 10:08:43 -0500

berak gravatar image

updated 2015-08-19 10:36:40 -0500

  • "each with 1 photo" - your plan is doomed. you need like a good dozen (better two) per person.
  • "If we take a database of 100,000 users ... how long will it take for the results to return?" - a plain 1-nearest neighbour search is applied for all of opncv's algos here, so: time_for_getting_the_features + 100,000 x time_for_norm_or_compareHist (no worries, that's still fast. training on 100,000 images will set you back for years (or will just explode) when using a 'global' model (like fisher or eigen) an average pc hits the memory/cpu contraints for a pca at ~ 20k 100x100 images (lbph is far less restricted wrt. image count))
  • "...would be a quad core" - unfortunately, not much of it is optimized for multicore (norm(), compareHist() or such are, but not the recognition process itself).
edit flag offensive delete link more


@berak, about the memory constraints, I am not usre that 16GB RAM is average? :D

StevenPuttemans gravatar imageStevenPuttemans ( 2015-08-20 02:57:20 -0500 )edit

^^ oh, right. (my boxes are all notoriously low spec)

still, try to imagine taking the eigenvalues of a 100k x 10k Mat ;) wherever the actual limit is, there sure is one.

berak gravatar imageberak ( 2015-08-20 03:07:23 -0500 )edit

True there must be one ;)

StevenPuttemans gravatar imageStevenPuttemans ( 2015-08-20 04:17:25 -0500 )edit

Question Tools

1 follower


Asked: 2015-08-19 10:01:00 -0500

Seen: 229 times

Last updated: Aug 19 '15