AI Face Detection / Recognation NVR4216-i

lodeon1337

n3wb
Joined
May 26, 2020
Messages
16
Reaction score
5
Location
Germany
Hi, i have a problem with my new NVR4216-i.
I bought it to use fast face recognition, the face recognition of the cameras in my opinion is much worse and slower than that of the NVR.
I seem to have set everything right. AI by NVR, NVR is armed, the faces are also recognized very very quickly and have a box around them.

But when I later look at AI Search Face Detection or Face Recognation, it doesn't show anything there. 0 Results found?!

The recognized faces also do not appear in the live view with AI.
Only the IVS messages appear there, but no faces, although these are outlined in the live videos.

How and where are the faces recorded and what am I doing wrong?
The firmware is the latest.

Thank you very much, Michael
 
Last edited:

lodeon1337

n3wb
Joined
May 26, 2020
Messages
16
Reaction score
5
Location
Germany
Thank you for your answer bigredfish. :)
Yes, I tried that too. Unfortunately no improvement.

But I don't think that's actually necessary with AI via NVR, because cameras without FD should also work because the NVR does the detection.

I see in the live image that the face is recognized by the NVR (box above the face), but not in the FD search.
Under Alarm / Alarm Info the NVR also shows that a face has been detected.

Just not in the FD search or in Live View with AI on the right. I also believe the frimware is very buggy ....
 

ArnonZ

Pulling my weight
Joined
Feb 11, 2020
Messages
212
Reaction score
128
Location
here
I have the 5216-16P-I
Not sure about the performance yours, but the 4216 is not as strong. it also had a pretty bad review here.
First, the idea of face recognition in general is great and doing it using the NVR have a different advantage.
The 5216 too is not ready for market and suffers from many problems but more than everything - they have missed with the implementation of FR.
Not sure why you think that FR by the NVR is faster, especially since you didn't get the function to work.... Anyway, the reason it should be done by the NVR is usinf a common face DB for all system.
You want a central DB that knows the right faces and that each detection would check for a match in that DB: if cam 1 now knows Danny's face, all cams would identify him as Danny.
Currently, it's not even doing AI: you they not add more and more images of Danny to improve his identification, they simply add another person names Danny to the DB.
Google Picasa used to add more and more images of Danny to Danny's DB and improved his identification the more you added images of him. There is no "learning" here and it's not AI.
As I was saying, the implementation is currently BAD. What the FW is capable of, the SW does not do. The bad SW misses the whole point of FR and it's a shame:
In general, FR is not as reliable as other functions (that's fine with me) and therefore, had to be dealt in a total different way:
It had to function as an extra filter for SMD and/or IVS - If you have a line crossing, if it's human AND it was not a known person, set off the alarm.But it's fine to skip the alarm if it was 99% Danny.
But as a standalone, it's pretty much useless. Currently, it only works as a standalone feature.
The second problem: you may take many face captures of the same person (the camera is doing it fine). BUT, as long as you follow the same person that is still in the frame, they had to avoid making a decision for each and every capture: if three of my captures were not clear enough and one was, it is still Danny.... it's not three strangers and Danny!!!
The third problem is related to the other two (or eliminated if you fix the other two): As I was saying, the FR is not as reliable as you'd like it to be and that's fine with me. The distance, angle, light etc are limited and not every capture is good enough for FR.
If it can get a good FR of me 5m away, why starting when I'm 15m away? as a standalone feature, it's got no way to limit the area of detection as SMD and IVS are doing. If it was an extra filter for SMD or IVS as I suggested, that problem could be solved.
It's got a great potential. It could really bring lots of customers from BI to Dahua NVR. But at the moment, the bad SW implementation made it totally useless.
For now I'd suggest: don't buy it. It's a total wast of money.

If they ever fix it and turn the FR feature to be usable, I'd be the first one to make a nice Youtube video about it.

Back to your problem:

The green square notes facial detection. what made you think that the face RECOGNITION works in your care. detecting and identifying are two different things. it's very easy to detect face, way more difficult to identify.
If you do not see recognition in the AI search and not at the live view and probably your face DB is empty, what makes you think that it's working.
Please make sure that you've set it for face detection AND face recognition AND chose smart plan without the three of them it won't work. ALSO, you should avoid using human detection for the cams you want to use for FR. Human filtering can still be done VIA SMD and/or IVS but not the human detection function.
Also, your NVR supports less cams without build in FD for face recognition. Not sure but I think that only 1 or 2. If the cam itself support it, I believe it's up to 8 of the cams but even than, check if you set the NVR to do the FD or not.

@bigredfish :
It could work also with IVS smart plan if you're doing the both FD and FR by the NVR. in such case, you are limited in the amount of cams you may use for FR, but you can get both IVS and FR by the same camera.
Mind that you can choose if you want to use some of the features but the cam or by the NVR or both. if you do it by the NVR, the smart plan does not matter.
 

bigredfish

Known around here
Joined
Sep 5, 2016
Messages
17,454
Reaction score
48,660
Location
Floriduh
Thx. I knew about the number of cameras limitation but not having played with the “I” version NVR enough to have sorted out that you didn’t need it activated on the camera as well.

My very minimal testing on Recognition was done on what is now over a year old unit and honestly I gave up on it. But from your analysis it doesnt sound like they have it sorted out yet.

I think as I’ve seen in Dahua’s own demo videos that there are a very fine and specific set of parameters at very close distances with very precise angles needed to get good results. But the additive effect of AI identifying two or more different angles and captures and building that into a single person in the database really is the “trick” to real AI recognition and it does t sound like it’s working yet.

thanks for your explanation
 

ArnonZ

Pulling my weight
Joined
Feb 11, 2020
Messages
212
Reaction score
128
Location
here
I think as I’ve seen in Dahua’s own demo videos that there are a very fine and specific set of parameters at very close distances with very precise angles needed to get good results
Again, I never expected the FR to be reliable as standalone feature. you can never expect every auto capture to be good enough for FR.
But just because of that, I could expect it to be an additional filter for SMD or IVS. Could have been enough for me to filter some of the false positive alarms.
When my wife returns from airport at 3AM, she can even face the cam to enable a good capture and avoid setting off the alarm. If it worked half of the times, it could be good enough.

currently, I have a different Chinese solution for that problem: coronavirus. Since the beginning of it, she never returns from airport at 3AM.
I would still prefer a nicer solution form Dahua.
 

m00st

Young grasshopper
Joined
May 13, 2016
Messages
30
Reaction score
22
I have played with Face Recognition on my 4216-I some time ago and it is a nice toy, but not bullet proof in any way. The good thing is that I've not seen it do a false positive (falsely recognize a stranger as a known person). But it very often does not correctly detect a person that is known in the database. So for opening doors, it will work just don't expect the door to open 100% at the time.

One thing you may want to do is - assuming you got your unit from EmpireTech - reach out to Andy and ask for the most recent firmware. It's from mid-August if I am not mistaken. It added quite some features and in my case it seems it solved an annoying bug on the AI/IVS (which is the more practical reason to use this NVR): In my previous version it detected IVS tripwires/intrusions from camera A and B while someone was in the picture of camera A. Doesn't sound like a huge deal at first, but it is quite annoying. Fortunately, this got solved. Since I have IVS detection by the NVR on 3 camera's I can't rest face recognition but I wouldn't be surprised if that improved too.
 

ArnonZ

Pulling my weight
Joined
Feb 11, 2020
Messages
212
Reaction score
128
Location
here
The Face Recognition feature is not and cannot be "bullet proof". It's not expected to be as reliable as IVS or other methods.
The limitations of FR could and had to be dealt of with a good SW and that's (yet?) not the case.
As m00st was saying, the false positive of that technology is very rear. I've never had a Single false positive.
Saying that, the SW had to get the various face captures from the camera, follow the same object and make a single decision based on the best face capture.
In order to deal with problematic ID at long distance, it had to enable detection only in a zone inside of the field of view. The best way to do so was to merge it with IVS.
That way, it could have been (and had to be) an extra filter for IVS (not only movement in that square, a human movement and if it's not me or someone I know...)
At the moment, the FR is no more than a gimmick.
An Expensive Useless Gimmick (EUG).
It's a same cause it could really be a killer application - one thing that could justify choosing Dahua NVR over BI. It's not the case.

@m00st: There are some newer versions all the time. @EMPIRETECANDY is really great and sends each update really quickly.
But there are still no attempts from Dahua to step in the right direction.
 

m00st

Young grasshopper
Joined
May 13, 2016
Messages
30
Reaction score
22
Agreed, the face recognition (FR) is nothing to buy this NVR4216-I or any other Dahua AI NVR for. But the NVR4216-I is still a great contender if you have an installed base of camera (Dahua or other Onvif) or want to save on the camera's by not buying SMD versions.

I've been running a few camera's around my home for some years. Some Dahua (of various quality) and a few other brands. Whatever camera I've used so far (without Dahua SMD) is useless in reporting movement. More useless than the FR :) But since I use the 4216-I's smart AI feature I can pretty much rely on it again. And sometimes there just is no affordable Dahua option. Example: I recently bought a cheap (<35$) 170 degree overview camera that I use with the NVR's AI to direct my Dahua PTZ to the point of detection. Not aware of any affordable Dahua wide angle/fisheye that could do the same and in this case I get near flawless detection+alerting of people and vehicles. My Dahua PTZ's autotracking will take care of some up-close shots ;).

As for the quality of the AI/SMD logic in the NVR: False detections are rare (an occasional phantom, but no more birds/cats etc), and if properly configured (does take some tweaking and trying) you'll catch >90% of humans or vehicles. I have one SMD enabled camera (1A404XB-GNR from Andy) and the NVR's logic seems to beat that of the camera. I previously relied on a PIR+Microwave detector to detect presence at my front door, but the NVR solution is superior.

So this NVR4216-I and maybe even more so the PoE versions (NVR4208-8P-I or NVR4216-16P-I) are great options if you're starting from scratch and need a typical (PoE) setup for protecting a property. One of those NVRs and a few simple Dahua/Imou camera's are a dirt cheap and reliable solution with human/vehicle detection (4 camera's as long as you don't use FR).
 

ArnonZ

Pulling my weight
Joined
Feb 11, 2020
Messages
212
Reaction score
128
Location
here
I'm not sure if the 4kS2 NRVs are not supporting SMD.
The power of the I series is the FR, it cost much more and currently, the Dahua implementation makes it completely useless.
I can only hope that Dahua would make a big revision in the whole concept to make that feature practical. I doubt it.
 

m00st

Young grasshopper
Joined
May 13, 2016
Messages
30
Reaction score
22
Up until I received a new firmware update from @EMPIRETECANDY i thought SMD and the IVS human/vehicle detection were the same. But in the recent update for my 4216-I SMD became a distinct configurable/ searchable option under the AI menu.

For now, the 4 camera smart AI/IVS (with human/vehicle recognition) suffices but the next series of AI NVRs is unattractive in pricing because of PoE ports I don’t need. If anyone (@EMPIRETECANDY , @bigredfish ?) has insights to what level the 4KS2 NVRs support the smart AI features of the 4.0 software (including ANPR) that would be very helpful!
 
Top