Hikvision Facial Recognition Demo From DVS

Consider how often a person is missed.

It's understandable that in demos, the focus is on showing correct matches. But, in production, if you are trying to alert against someone or even find where they were at a certain day or time, the number of misses makes a big difference.

The worry is that if an expectation is set that "we'll get this person if he comes back" and the system misses, it will create an upset user. 

Counting misses is harder, simply because by definition they are not flagged but checking the ground truth will help give a better sense of performance.

Beware Stranger Comparison Alert

DVS shows off an interesting feature where the system will notify you if a 'stranger' is found, i.e., someone who does not match anyone on the known person's list.

Even if you have a closed population, there are going to times when a person is not matched (technically the face is below a similarity level set) because the person's face was obscured or in a shadow or has a hat on, etc.

Unless you are in a very high threat environment, you will probably want to not use something like that to avoid ongoing nuisance alerts (i.e., the stranger just turns out to be 'Bob' in bad lighting).

False Matches Don't Show In Small Offices

The DVS demo shows no false matches (specifically a person being matched to someone else at a similarity level of even 70). That's believable. The bigger challenges comes with more people (e.g., a mall or a bank branch or a public place). In such environments, 1000s of faces a day will be ordinary but will make matching more complex.

If you only plan to deploy face recognition in small offices, then that's not a worry.

Multiple Matches Per Fatch

One interesting thing is that the DVS system regularly captured multiple matches per face in a single scene, e.g. 4 matches, same person, 2 seconds:

This becomes a usability problem because you have 2-3x the number of faces shown as actual people past. Also, it can increase matching issues (e.g., false positives) because each of those 4 face captures will match differently.

Capturing Harsh Angles And Occluded Faces

One other interesting thing is how tolerant the system is to capturing occluded faces and harsh angles, e.g.:

I am curious to see how well those captures do with matching (e.g., the woman in the center left is going to be hard to match accurately).

Hikvision itself has a China demo video where they have similar matching approach:

Here is a clip from that video:

Our Testing To Come

Unlike the UK, in the US, Hikvision has not officially released facial recognition (specifically their facial recognition camera) so we will wait for that before doing a test.

John, good points all. I'd like to raise the point that the size of the database will have a lot to do with matching accuracy. So a 99% match to Dave is great. Especially with the different lighting (Near IR, sunlight, etc.). But exactly how many people, and how many images per person, was Dave's probe image compared to? A few hundred to a few thousand people would make the demo more interesting and relevant to get a handle on accuracy (FA & FR) and response time. Finally, I think the interface is pretty good! 

That is an outstanding demo. Much better than the Roadshow demo, tbh. Real world results may vary, but that shows the capabilities well.

Just amazing! Incredibly good trained neural network. What is the price of such NVR and how many channels supported ?

Hi

I am waiting to see when Hikvision releases it in the US market. A mature, exigent and extremely competitive market; one can easily replace "A" with "The most".

I would surmise that this require mroe processing power from the DVR/NVR than I have yet seen. Perhaps there is special DSP chip inside for this purpose. It would seem to me that a powerful PC( new I7 Chips, lot of RAM and Storage) would be more suited for this. The best solution may involve cloud computing so that the amalytics are performed in a data center . We all know how we all feel about that part:  a cloud system operated by a Government ...

I am ambivalent about all of this... We all seem to grow more ambivalent about the reach of current available technology.

Here is Dahua's face recogntion demo video:

They use a sketch of a person to do matching.

The main question / concern is how many other people are in the system. That's what makes facial recognition challenging. If you only have a handful of people captured, it's pretty easy to find the person you want.

Here is an excerpt of their results screen:

Clearly, they imported the same images over and over. The sketch matches 83% to a mugshot photo. Whether that is good or bad depends on how many other people would match over 83% in a real scenario with thousands of people captured.