5.4.6 - May 17, 2021

[m_listed]

There's a bug where alerts don't recognize faces starting with capital letters. So "face1" and "face*" will work, but not "Face1" nor "Face*". Also, another bug where renaming faces to change capitalization doesn't actually register in the system as a change, so e.g. renaming "face1" to "Face1" will seem like it worked, but it'll still be "face1" in the system.

 

[Dave Lonsdale]

I'd written support last week about some suggestions for the AI and I don't know if Ken got the inspiration himself or from someone else or for me, but this is exactly what I had suggested. Theoretically fixes problem like "Person approaching camera, still too far to see his face, but DeepStack stops detecting as soon as it detects a person and ignores his face as he gets closer to the camera." It should be able to keep detecting (I put "zebra" in my "to cancel" box so it'll keep going) until it maximizes the confidences involved.

At the same time, makes me fear for the future. This is a huge step towards detecting anything and anyone at all times, now in the hands of all of us xD

@Dave Looking good! I like your zoning. I have similar zones for one of my cameras. But I have a background zone that covers the entire frame so that Blue Iris will keep tracking objects in between zones and/or have an easier time understanding what constitutes an object when they're in-between zones. Did you have any missed motion that you know of? If not, then maybe that background all-encompassing zone isn't that important to have.

Hello m_listed
When I said I had temporarily inverted Zone A (the full image), this was only whilst I took a snapshot for the post so that the other zones could be seen. Sorry that wasn’t clear enough. But for large moving objects eg cars that straddle the “Object crosses zones” I don’t think having Zone A makes any difference. I have not been able (yet) though, to find a setting that stops DeepStack from cancelling some car alerts that BI has triggered perfectly.
An additional point:
Having looked at the confirmed alerts more carefully, I have only today noticed that even when Zone A does not appear in “Object crosses zones”, if “Not required for re-triggers” is checked, BI maintains its triggered state even when only Zone A has motion. So with trees blowing in the wind round the periphery, the triggered state can be maintained for minutes instead of seconds. I’ve been a Blue Iris user since 2013; it’s high time I knuckled down and learned a lot more.

 

[m_listed]

There is another bug as of 5.4.6.2 when creating Faces. Often, DeepStack won’t be able to find a face in your Face images. Even the ones that it spits out itself into your set Unknown Faces folder. But when you add a Face image and it fails to find a face in it and thus add it to the DeepStack database, Blue Iris doesn’t understand its refusal and acts like your Face addition was successful.

I didn’t realize this until I ran deepstack manager and saw that only 6 of my 20 faces were actually registered! And I was wondering why it would always be finding some of faces and not others.

So yeah, I’d check on that. Looking at the DS API shows that you can register multiple face images under a single name. I wonder why Blue Iris creates a new name for every face instead.

 

[Dave Lonsdale]

Perhaps for the camera I've featured earlier in this thread, there is no ideal setting for DeepStack to confirm all vehicles passing the end of my drive:-
When running the BVR file through Analyse with DeepStack, the truck is flagged in red for 1200mS. However, with live data, 200mS of this is taken waiting for the zone crossing plus my (short) Make time setting of 200mS. The Blue Iris trigger/alert mechanism has no problem with this because a perfect alert image is created using the preceding key frame aggregated together with the subsequent Pframes. I'm guessing that this perfect image is not or cannot be sent to DeepStack and BI must wait for the next individual key frame to send. By this time, in this example, the truck has gone.

It's a shame that, unlike Hikvision, this Dahua 5422-Z4E on max zoom has a minimum key frame interval setting of one second. I think an iframe interval of 0.5s would solve it.

If BI can only send key frames to DS, could this be a similar to others with cancelled alerts - when using a more normal Make time of one second, it could actually be up to two seconds before Deepstack gets its first image when, for example, the wanted face has turned?

Could someone else more expert than me please advise?