Deepstack does no Face recognition, only "Person"

Post Reply
Grisen
Posts: 2
Joined: Wed Sep 22, 2021 11:39 am

Deepstack does no Face recognition, only "Person"

Post by Grisen » Thu Sep 23, 2021 1:13 pm

I have Deepstack up and running with person detection working fine. However, only occasionally the face recognition is triggered. This results in that I get alerts with "personx%", rather that "personx%, uknown/<face_id>". I guess this could have to do with that an alert is fired as soon as any of the conditions ("person") is triggered, rather that when both criteria "person", "unknown" are met. When I re-run the clip, and under "Test and tuning" select "Analyze with Deepstack", I get the expected rectangles. I have tried to modify my camera settings by disabling "Begin analysis with motion-leading image", set "+real time images" = "5" and "analyze one each" = "2sec", but without success.

Should I lower the motion sensor sensitivity of the camera to increase the chances that the image that is sent to Deepstack for analysis includes both "person" and "face"? Any other settings I could tweak to make this work?

Any help is very much appreciated.
User avatar
YrbkMgr
Posts: 106
Joined: Sun Nov 24, 2019 12:56 am

Re: Deepstack does no Face recognition, only "Person"

Post by YrbkMgr » Fri Sep 24, 2021 5:55 pm

This is my general approach with DeepStack object identification issues. I'm not an expert, so your mileage may vary. I'm assuming that the AI tab of Global Settings has face rec enabled.

Generall7y speaking, if T&T/Analyze with Deepstack performs properly, the focus should be on camera settings and/or stream optimization. You might consider reviewing the Onboarding section of the forum for camera stream optimization. There is also a Wiki in another forum that details optimizing BI and a substream guide.

Does the problem happen on more than one camera? If there's only one camera or only happens on one camera you should focus on determining if DeepStack is getting data fast enough to perform object identification when the camera is triggered. To do that, you can review the DS "log" -> Status Window -> Deepstack Tab. If you're not familiar with the DeepStack Tab of the Status window, this may help: Canceled Alert with 91% deepstack.

Do the motion rectangles from DS line up with the triggering objects? That will reveal possible camera optimization/stream issues. If it looks like it's matching up, then it's buried in the settings. If they're not matching up, focus first on stream optimization.

If you post your Trigger settings and AI settings that'll get you more help here.

It's a bit of trial an error for those who don't have "Zen" with BI/DS, but the above is how I approach it so maybe that helps.
What I love about standards is that there are so many to choose from!
Grisen
Posts: 2
Joined: Wed Sep 22, 2021 11:39 am

Re: Deepstack does no Face recognition, only "Person"

Post by Grisen » Mon Sep 27, 2021 8:43 am

Since I wrote the first post I experimented a little with my cameras and set up a new trigger for my "Home mode"-profile. I used the same motion sensitivities etc. but changed the Deepstack-setting "To confirm" from "person, unknown" to only "unknown". Now I get a lot of "unknown" triggers (all cameras) in my Alert-list so it seems the face recognition somehow works given the placement of the cameras etc. Not sure if that helps but I think this might indicate that the camera is triggered before Deepstack gets the chance to analyze an image with a face (opening the door scenario). Or am I reading this wrong?

Regarding the rectangles under T&T they are slightly off. For example, the rectangle with caption "unknown" is placed where the face was on the screen 0.2-0.5 sec ago. Is this lag acceptable or a sign of that I should work on the stream optimization that you mentioned?
trigger.JPG
trigger.JPG (64.49 KiB) Viewed 340 times
AI-settings.JPG
AI-settings.JPG (88.63 KiB) Viewed 340 times
Post Reply