5.4.1 - April 3, 2021 - Support for DeepStack facial recognition

So my deepstack setup works great. I have blue iris set up for someone else. The last 2 days the deepstack ai has been working fine then around 1am deepstack is reported in the bi log as not running or unreachable. Anyone have any idea why.
 
I gave it a good go all last night and most of today tweaking but was forced to go back to a older build before an update was changed to the Motion detection. Unfortunately, the updated motion detection is producing constant triggers that it was not before. From what I can see the rectangle is getting bigger as an object moves off to the edge of the camera's view which is then crossing a zone. This is happening with people that walk past the house in which never before would trigger off the camera. I tried all settings to stop this but the problem is the rectangle just suddenly jumps bigger as someone walks by. The newly updated motion detection is causing havoc which is a shame as I was getting nice results with Deepstack. Deepstack will of course trigger a confirmed alarm as its people I want to detect, just not people triggering outside my set zones. I cant edit the zones as this would cause other issues. There was a clear line between the path and my drive way but the rectangle of a human only at certain points is almost double the size. Cant rectify this is making the Max size cancel as I need it the way it is for people getting closer to camera.

Dan
 
  • Like
Reactions: beepsilver
Is it possible to train the facial detection with multiple pictures with the same label? I tried last night but it created some errors for me

To add on to this, I am not sure how to confirm facial recognition is working or not. I don't seem to be able to cancel on faces, and without any memo or something I can't confirm what it's doing. Am I missing something obvious?
 
  • Like
Reactions: m_listed
To add on to this, I am not sure how to confirm facial recognition is working or not. I don't seem to be able to cancel on faces, and without any memo or something I can't confirm what it's doing. Am I missing something obvious?

Until analysis of .bvr files is supported all you can do is analyze a .jpg file. To do this, preview a bvr alert file from the alerts view and pause it when your target is in clear view. Take a snapshot (upper left camera icon in the video frame), then save it in the default location. Next, doubleclick the newly created snapshot in the alerts viewer and your static .jpg file is displayed. Right click on the image and select 'analyze image with deepstack.'

At this point it doesn't seem BI allows you to save multiple .jpgs of different faces for the same person with the same name unless something has changed in the last couple of BI updates. For now, I've simply named each photo like this: dave dave1 dave2
 
Seems that I’m getting it to grab a lot of faces and put in the new faces folder I have setup. Unsure how to get it to work to actually match for footage. Tried putting the names into any of the detection boxes but never get any hits
 
  • Like
Reactions: m_listed
Question - I am really wanting to use the new AI facial recognition to trigger alerts when it sees specific people (restraining order issues, stalkers, etc,). Is there a way to set this up in BI with the new features?
 
Question - I am really wanting to use the new AI facial recognition to trigger alerts when it sees specific people (restraining order issues, stalkers, etc,). Is there a way to set this up in BI with the new features?

Yes there is. Of course YMMV on the accuracy based on your camera heights, amount of light, etc.
 
I have 5.4.3.8 setup and working well with Deepstack. I get alerts of 'person, car, etc' upon triggering with emails. I have any of your had any success with the facedetection? I have uploaded multiple pictures from different angles of myself, and then put the names of the pictures in the cancel line but still get alerts for person detection. Am I missing somethign, I wasn't able to attend the webinar to learn more about the face detection capabilities.
 
I have 5.4.3.8 setup and working well with Deepstack. I get alerts of 'person, car, etc' upon triggering with emails. I have any of your had any success with the facedetection? I have uploaded multiple pictures from different angles of myself, and then put the names of the pictures in the cancel line but still get alerts for person detection. Am I missing somethign, I wasn't able to attend the webinar to learn more about the face detection capabilities.

Unless I missed a post in this thread, I have yet to see anyone having success on the face part. I assume it still work in progress. Like you, I have had excellent results in the object detection part.
 
Despite not being up to date on DeepStack facial recognition, I think the fundamental flaw has already been pointed out - you choose a high resolution camera for a particular situation to identify a person based on its DORI performance and then use its substream having a 10x lower resolution for facial recognition. Not surprising it struggles - or maybe my knowledge is too far below where it needs to be to understand.
 
  • Like
Reactions: MicahJames
Despite not being up to date on DeepStack facial recognition, I think the fundamental flaw has already been pointed out - you choose a high resolution camera for a particular situation to identify a person based on its DORI performance and then use its substream having a 10x lower resolution for facial recognition. Not surprising it struggles - or maybe my knowledge is too far below where it needs to be to understand.

In my case I have it setup with a single 8MP camera that isn't setup with a lower resolution substream and is just continuously recording. I have gone as far as staring into the camera to see if works and I suspect it is not fully functioning. Additionally the pictures I uploaded include high res and medium grade photos. Will let others know if I learn anything else from tinkering with the face detection.
 
So just to confirm, no one has had Faces work for any kind of notifications, labeling, etc. right? Objects work, but Faces don’t.
 
I used it in the: "To cancel" field. I added photos of myself and the wife. It worked but also thought my neighbor was me (Profile did kinda look the same). I deleted the "To cancel" entries because I don't want to miss someone it thinks looks like me.
 
I used it in the: "To cancel" field. I added photos of myself and the wife. It worked but also thought my neighbor was me (Profile did kinda look the same). I deleted the "To cancel" entries because I don't want to miss someone it thinks looks like me.
How did you figure it canceled because of (what it thought was) your face? Could it have been something else, like it never detected anything? Because I don’t see any confirmation in BI that Faces are being taken into consideration whatsoever. No alert image burn-ins, no memo labels, no logs, no alerts.
 
How did you figure it canceled because of (what it thought was) your face? Could it have been something else, like it never detected anything? Because I don’t see any confirmation in BI that Faces are being taken into consideration whatsoever. No alert image burn-ins, no memo labels, no logs, no alerts.

I knew the neighbor had been over (he picked up an Amazon package for me) but I didn't receive an alert. I checked to see why and I found the alert snapshot and it identified him as one of the pics of me I entered in the "To cancel" field. Then I compared the clip of him at my front door and could see the resemblance. That alert snapshot has been overwritten but here is one of me identified properly (alert canceled). I still use the "To cancel" feature for my inside cameras.
 

Attachments

  • AMA.20210503_164736.0.3.jpg
    AMA.20210503_164736.0.3.jpg
    38.7 KB · Views: 87
  • Like
Reactions: sebastiantombs
I knew the neighbor had been over (he picked up an Amazon package for me) but I didn't receive an alert. I checked to see why and I found the alert snapshot and it identified him as one of the pics of me I entered in the "To cancel" field. Then I compared the clip of him at my front door and could see the resemblance. That alert snapshot has been overwritten but here is one of me identified properly (alert canceled). I still use the "To cancel" feature for my inside cameras.

That’s amazing, I didn’t know it labeled faces like that, only objects. Anything special you did to get it to do that? I used to get some “unknown face” snapshots, so I added the known faces and stopped getting the unknown face snapshots... but without the face alert image labeling or face alerts. :(
 
I am not getting good results as it seams to recognise Alloy Wheels as faces and 80% are captures of car alloys like some sort of fetish :)