Face Recognition Algorithm queries

[GProf]

Hi,

I'm playing with the SDK and the Face Recognition Algorithm.

So far what my simple tests work great (registering/unregistering a user, saving/loading a database).
The documentation and currently available article are helpful in getting started, 
but it doesn't cover the algorithm used itself in greater detail. 
I imagine sharing algorithm insights is tricky in terms of IP, but I have some queries/concerns that hopefully can be covered:

1. ​Is the Face Recognition Algorithm using both RGB and Depth cameras or just one (if so which one) ? The main motivation is trying to work out the ideal scenario to get most accuracy and avoid problematic setups (bad lighting, distance: what is the maximum distance a user should be to be accurately recognized ?)
2. What is the maximum number of faces that can be recognized at the same time ? (e.g. a group photo scenario)
3. What is the maximum number of faces that can be stored in a database ?
4. Do the queries slow down if the database gets really big ? (If so, at what number of entries roughly does it start to get slow enough to cause delays and what strategies do you recommend on tackling this ?)
5. What other hints/tips not covered in the documentation do you recommend for optimal facial recognition in terms of accuracy and speed or retrieval ?

Thank you,
George

P.S. I'm currently using RSSDK 2015 R4

[Xusheng_L_Intel · Accepted solution]

1, currently we only use RGB camera for face recognition.

2, the max number of faces can be recognized at the same time is 4.

3, no more than 33 faces can be stored in the database

4, big database will affect the performance

5, please clarify your question.

Thanks!

[GProf]

Hi David,

Thank you very much for the helpful response.

Regarding question 5: it's a more general/open question.
I was wondering if you could share any useful advice from your experience working with the algorithm
on what data produces better results (as opposed to setups that produce bad or no results).
For example avoiding very poor lighting (lights too dim/flashing/etc.), having user in a certain distance threshold.
What factors potentially affect the algorithm in a bad way (so I can avoid these) ?

Thank you,
George

[PKusm]

I have just read the release note of RSSDK R5, and there should be an improvement on the Face Recognition algorithm.
However, it is stated on the R5 release note that the Face DB can only support up to 20 faces.
Regarding your question number 5, It is also worth knowing that the face registered to DB should be facing towards the camera. Registering face in various pose will reduce the accuracy.
Right now, I am struggling to make an automatic face registration method so that users do not have to push any button to register their face. And, it is quite tricky to guarantee that user's face is always registered while facing towards the camera.
Anyway, I am upgrading my RSSDK to R5 soon, and I will inform you if there is something useful.

Cheers!

[GProf]

Hi Putu,

Thank you very much for the update.

I see what you mean with registering faces, and I agree, font facing is better.
Also, the automatic option sounds like a tricky one (ignoring blurry frames is person is moving, etc.)

To get past the DB limitation, is it currently possible to swap between DB buffers at runtime ?
If so, may I direct you to this topic ?
If not, is that a feature that could be added in R5 ?

Thank you,
George