Face Recognition Driver Now Available!

[Shivam Paw]

So if ANPR wasn't enough for you maybe our new face recognition driver will wow you!

This driver makes use of Microsoft's Face API and allows you to, via a web interface, add the name of people and add faces to those names. The driver will then work similarly to our ANPR driver and once a programming action is triggered (in C4 programming) we will capture the image from any URL or IP Camera (supporting Basic & Digest authentication) and then try and identify a person in it.

If a face is identified then either the Known Face Detected event will be fired which will allow you to make use of the Person Name and Confidence Level (0-100) variables to program based on who was detected. If the face identified is unknown the Unknown Face Detected event will be fired and allow you to program based of that scenario.

The driver also includes the ability use an image in the push notification. The image available will be the image when there was last an identified face. This lets you, for example, have a push notification sent to you with the image taken when there is an unknown person detected.

Images

The Microsoft Face API provides 30,000 free requests per month. Each identification takes between 1 and 3 requests. Requests are also used when adding people and their faces. Generally, the 30,000 is enough for most installs if the trigger for detecting faces is reasonable – i.e. whenever motion is detected may not be reasonable, but when a line is crossed (Hikvision cameras or active beam sensor) would be a better use of the limit.

The driver is available on BlackWire Designs here: https://www.blackwiredesigns.com/store/control4-face-recognition/

 

[msgreenf]

Wow! You are killing it!

[Gary Leeds UK]

Nice Work 

[poseidonsystems]

I got to give this a shot 

[ILoveControl]

Spot on man! Glad to see you using the best AI service  

[zaphod]

What's the accuracy of the MS Face API?  Any idea on the rate of Type I and Type II errors?

[ILoveControl]

What's the accuracy of the MS Face API?  Any idea on the rate of Type I and Type II errors?

Very good question! I *think* its described in the API documentation, depending on how the API is called and what is allowed to be stored it does get “smarter” on common images you keep posting - if you choose to use the learning feature. This way the ML model used for the calls starts learning from the false positives etc.


Sent from my iPhone using Tapatalk Pro

[Shivam Paw]

3 hours ago, zaphod said:

What's the accuracy of the MS Face API?  Any idea on the rate of Type I and Type II errors?

 

3 hours ago, ILoveControl said:

 

Very good question! I *think* its described in the API documentation, depending on how the API is called and what is allowed to be stored it does get “smarter” on common images you keep posting - if you choose to use the learning feature. This way the ML model used for the calls starts learning from the false positives etc.


Sent from my iPhone using Tapatalk Pro

 

This document by Microsoft provides very detailed information about the accuracy, how it works and ways to increase it

https://azure.microsoft.com/mediahandler/files/resourcefiles/transparency-note-azure-cognitive-services-face-api/Face API Transparency Note (March 2019).pdf

[zaphod]

Interesting info, but that doc doesn't really go into much hard detail - the answer is more "It depends."

 

Quote

How accurate is Face API?

Because Face API is a building block for creating a facial recognition system to which other building blocks must be added, it is not possible to provide a universally applicable estimate of accuracy for the actual system you are planning to deploy. Companies may share accuracy as measured by public benchmark competitions, but these accuracies depend on details of the benchmark and therefore won’t be the same as the accuracy of a deployed system. Ultimately, system accuracy depends on a number of factors, including the technology and how it is configured, environmental conditions, the use case for the system, how people to be recognized interact with the camera, and how people interpret the system’s output. The following section is intended to help you understand key concepts that describe accuracy in the context of a facial recognition system. With that understanding, we then describe system design choices and how they influence accuracy.

 

[Shivam Paw]

3 hours ago, zaphod said:

Interesting info, but that doc doesn't really go into much hard detail - the answer is more "It depends."

 

 

Yep, but it does give information on what you can do to make it more accurate

[ILoveControl]

@Shivam Paw I think with both your drivers plate and face what would be most useful to the dealers / users if you would provide the debug info of the success and or fail of the AI (I think you already do this still on holiday will pick up the drivers when I am back and add them to my project to play around with them).

The reason I say provide as much info as possible (I would just make them variables what I mean by variables, success / failure rate (think you do this), confidence level (think you do this)), you can program some "help" information.

Example for me what I would do:

1. Using Confidence level - If I allowed a person or plate in based on a low confidence level say 85% I would still set a programming event to notify me (email or notification) it worked but there were some issues you should have a look a the captured image and go to link ABC to see how better to improve - could be cam resolution, position, light exposure etc.

2. Same as above for success / failure rate.

Question:

1. I think the Plate driver attaches a snapshot url from the camera to the notification (if I saw the image right) - is this correct?
2. Can  you enable snapshot URL for email? Use case: Based on my suggestions above I could then store the images to see what angel etc is right and wrong to adjust for better clarity.

 

[Shivam Paw]

4 hours ago, ILoveControl said:

@Shivam Paw I think with both your drivers plate and face what would be most useful to the dealers / users if you would provide the debug info of the success and or fail of the AI (I think you already do this still on holiday will pick up the drivers when I am back and add them to my project to play around with them).

The reason I say provide as much info as possible (I would just make them variables what I mean by variables, success / failure rate (think you do this), confidence level (think you do this)), you can program some "help" information.

Example for me what I would do:

1. Using Confidence level - If I allowed a person or plate in based on a low confidence level say 85% I would still set a programming event to notify me (email or notification) it worked but there were some issues you should have a look a the captured image and go to link ABC to see how better to improve - could be cam resolution, position, light exposure etc.

2. Same as above for success / failure rate.

Question:

1. I think the Plate driver attaches a snapshot url from the camera to the notification (if I saw the image right) - is this correct?
2. Can  you enable snapshot URL for email? Use case: Based on my suggestions above I could then store the images to see what angel etc is right and wrong to adjust for better clarity.

 

I’m not sure what you mean by providing success and failure rates. The driver does have a confidence level variable which can be used for your 1st example. 
 

as for your questions. 1) yes that’s right. 
2) I don’t think it’s possible in C4 to attach an image to an email (without other drivers). Most IP camera drivers do let you send a snapshot via email though so you can try that. Failing that, a push notification image would work. 

[rolldog]

I know this post is kind of old, but I ran across it while looking to see if QNAP's QVR Face AI used with their QVR Pro Surveillance System could somehow be used to create triggers in Control4. Right now, I have all my cameras setup using my QNAP as an NVR, but the only Control4 integration I can find for QNAP is as a network storage device, nothing else. It would be kinda cool if someone was approaching your front door, you could set things up like for the front door to lock/unlock depending on who it is.