Unrealistic Face Recognition Demos Where Everyone Looks Right At The Camera

JH
John Honovich
•Oct 28, 2018
IPVM

Take a look at this video, notice that everyone looks directly at the camera:

It is not clear how high the cameras are mounted, they could very well be 4' high as they are at Megvii's offices and/or it could be that subjects are told to look directly at the cameras.

The challenge is most people do not look directly at surveillance cameras due to a combination of (1) cameras not being normally mounted 4' high (10' or higher is more common) and (2) people having a tendency to look down when they walk and (3) people not walking directly in line to the camera (either the camera is tilted left or right or the person is not walking perpendicular to the camera, etc.).

The result is a good marketing video but an unrealistic portrayal of real use. If every camera can pick up every person, you can graph the route and timing of every person. The problem is that, regardless of the vendor, it will be logistically quite hard to accomplish this due to camera positioning, actions of people and the layout of buildings and streets. Once users realize that it generally does not work, it gets ignored. See poorly working video analytics.

(6)
(1)
Avatar
Simone de Titta
•Oct 28, 2018

Holy words, John!!!!... Holy words!!!!

Who knows if perhaps someday the market will see the light and understands finally that video analytics CAN’T be a product for supermarkets??...

Would you be ok if one day you went to a doctor for a health problem and he would get to you like “no, I don’t even need to visit you, get those pills, they are scanning you all and automatically understanding the problem and learning from your body how to solve it!!...”.. I’d not think you would love and trust that doctor that much?.. Or would you??...

None concerns and get amazed if for a medical problem in general the doctor visits you, makes specific analysis for your problem and prescribes those special pills, to be assumed in that way, in that hour and with that specific dosage, because it is you and you have that specific problem, in your specific age and living in these specific conditions... Or am I wrong??..

Of course there are cases of standard flu, or broken nail where you can actually go to get a VitaminC or a cream by the grocery, without the need of a special doctor; that’s ok... But when you are out of this standards (and in analytics this happens 90% of times!!!...), perhaps......

Of course miraculous cheap medicines with a great marketing can exist and can be sold at supermarket.. It’s legitimate in this stressed finance-based market.... On the other hand, of course it should not be a surprise then if you discover they work mediocrely and that out of the shown YouTube video’s conditions, they don’t work even praying in ancient aramaic the way they claimed......

Of course everybody care about their health.. Who cares about analytics?... Well, I think actually analytics made up a kind of headache epidemy, isn’t it? So perhaps it’s time to start to go for real doctors instead of nice brilliant shamans?...;))

Cheers,

Simone

 

(3)
UM
Undisclosed Manufacturer #1
•Oct 29, 2018

This was probably the most compelling element of the AnyVision demo I saw. The matching capability in cases of partially obscured faces and extreme angles. 

We spent some time live testing the thresholds of of detection, vs reference images pulled down that day from our social media accounts, and it was pretty solid.

Source was a random Axis 2MP indoor model on a tripod in the corner of the conference room.

(2)
(1)
MM
Michael Miller
•Oct 29, 2018
In reply to Undisclosed Manufacturer #1

I have been impressed with Anyvision.  We loaded up some video of a hallway which was packed with people from a 5MP camera at a stadium.  You couldn't fit another person in the hallway and Anyvision was picking out 40-60 faces at a time as people walked towards the camera which was mounted about 12ft up but we have a 22MM lens on the camera looking down the hallway. It is impressive to watch. 

(1)
(2)
U
Undisclosed #2
•Oct 29, 2018

Yeah, someone brought up NEC's "most advanced facial recogntion" the other day. 

It's like they are posing for a driver's license photo. 

NEC Demos World-Renowned Facial Recognition Technology at LPGA ANA Inspiration

Now do 10ft high, overlooking 25 people at various distances, in changing light conditions.

(1)
SD
Shannon Davis
•Oct 29, 2018
IPVMU Certified

This is just like relying on the demo videos camera manufacturers put on their website. I am huge Axis fan but sometimes the videos they upload I swear have been enhanced or have just the most perfect lighting conditions. 

(1)
Avatar
Clint Hays
•Oct 29, 2018
In reply to Shannon Davis

That's why I think a proof of concept test is worth way more than marketing material. 

(1)
JH
John Honovich
•Oct 29, 2018
IPVM
In reply to Shannon Davis

Most notable Axis camera demo: Axis: Take Down This Deceptive Commercial

Generally speaking, analytic overmarketing has more negative repercussions than camera ones. People seem to be more tolerant of imperfect video quality than imperfect analytics as the errors with analytics tend to cause more of an ongoing nuisance.

Avatar
Kevin Mundy
•Oct 30, 2018
Stanford University

John, I remember we drank the Kool-Aid back in the day too :-)  I learned early on how to sell "Facial Surveillance" and not Facial Recognition.  There is a BIG DIFFERNECE kids, do your homework well before class!

AnyVision is not just an analytic, its AI that learns as it goes and improves accuracy, IMO its the only way to go from here on out.

(1)
(1)
JH
John Honovich
•Oct 30, 2018
IPVM
In reply to Kevin Mundy

I remember us sobering fairly quickly though you were capable of taking the chaos longer than I was!

One thing I learned from 3VR was that when you 'optimize' rig demos all the time, you start to believe that's reality and it damages your ability to see clearly how real-world customers will use it. The funniest excuse that comes out of it (which still happens today, see comments of the Dahua face recognition test) is that the burden is on the user to change their cameras, behaviors, layout, etc.

Avatar
Sean Nelson
•Oct 30, 2018
Nelly's Security

Am I the only one who has become numb about Face Recognition, AI, Analytics, etc because of stufff like this? I dont even get excited anymore about this at all. Im glad progress is being made but its too bleeding edge right now and most of it wont work up to customers expectations.

I still feel more progress needs to be made to improve simplicity and use-ability of cameras. What the chinese need to focus on is:
- remove your darn nasty cable whips
- simplify and improve your software
- if you are going to make a web interface, dont make it IE only. If thats the only option, then dont even make a web interface.

You have unfinished business before trying to tackle facial recognition

(1)
JH
John Honovich
•Oct 30, 2018
IPVM
In reply to Sean Nelson

Sean, the other problem with AI, for your position, is that tech support calls would go through the roof. 

Things sold through distribution need to be made as simple and bullet proof as possible. While I am not saying Hikua has done this overall, cameras and recorders need far less attention / 'optimization' than AI.

(1)
UM
Undisclosed Manufacturer #3
•Oct 30, 2018

What about face recognition for access control, as an additional form of verification to a card, where employees could be told to look straight into a camera. Is face recognition used for that?

Avatar
Kevin Mundy
•Oct 30, 2018
Stanford University
In reply to Undisclosed Manufacturer #3

Yes you can use Face Analytics if you make sure you design it correctly.  Here are just a few of the things that make that easier in a "Controlled Environment".   Often you find these types of applications in Critical Locations such as Data Centers and other high security locations requiring additional analytics for single or multi-factor authentication.

- Cooperative Subjects who are often static when the images are taken making it 1000X easier for analytics to "find" the face (employee knows they need to look at camera for authentication)

- Enrolled Subjects/Template makes it easier to return persons whos Biometric Scores are higher confidence levels and give better accuracy in near real-time.

- Controlled Environments like Man Traps/Sally Ports where you can control light etc.

Take a minute and think about those and then you will understand why there is a big difference between Facial Surveillance and Facial Recognition.  In 90% of the applications you have Uncooperative Subjects in Uncontrolled Environments that are dynamically moving towards the camera through out the scene where the analytics are being performed. 

When you go to the DMV what do they do to ensure all of the images input are as similar as possible with the exception of ones facial characteristics?

- Cooperative Subject (stand on line, look forward, take hats and glasses off, no facial expressions)

- Controlled Environment (same light, exposure and backdrop for all images)

- All photos are taken at subjects "eye height" (camera is adjust so there is no vertical slope)

- All images typically have same pixel density making it easier to analyze and compare (resolution is same for all images)

 

(1)
(1)
New discussion

Ask questions and get answers to your physical security questions from IPVM team members and fellow subscribers.

Newest discussions