Test and Tune Motion and Deep Stack Review Orange Frame

General discussion about Blue Iris
iannecj
Posts: 3
Joined: Mon Aug 02, 2021 11:37 am

Test and Tune Motion and Deep Stack Review Orange Frame

Post by iannecj »

I am trying to troubleshoot why at times I am not seeing alerted or cancelled alerts for some motion events on a reolink 822A camera in the latest version of BI.

Basic question, if I do a review of a clip, I can overlay rectangles and motion from BVR recording which I do and I can run it though both motion and deep stack.

When checking motion with a sample clip 10-15 frames back I am seeing the moving element framed in yellow when it appears and eventually the entire view is framed briefly in orange.

I interpreted this as yellow around moving element as motion and frame orange around all as trigger.

Therefore when that happens I expect to have a cancelled motion alert or a confirmed alert but dont see either. In fact no event clip correspond unique to the time the ornage fram appear...not even close.

What could cause this?
User avatar
YrbkMgr
Posts: 587
Joined: Sun Nov 24, 2019 12:56 am
Location: Chicagoland

Re: Test and Tune Motion and Deep Stack Review Orange Frame

Post by YrbkMgr »

iannecj wrote: Mon Aug 02, 2021 10:08 pm In fact no event clip correspond unique to the time the ornage fram appear...not even close.
I was following you all the up to that sentence. Maybe a screenshot would help.
  • "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
  • If you know what your after, you'll recognize it when you see it.
iannecj
Posts: 3
Joined: Mon Aug 02, 2021 11:37 am

Re: Test and Tune Motion and Deep Stack Review Orange Frame

Post by iannecj »

YrbkMgr wrote: Tue Aug 03, 2021 12:45 am
iannecj wrote: Mon Aug 02, 2021 10:08 pm In fact no event clip correspond unique to the time the ornage fram appear...not even close.
I was following you all the up to that sentence. Maybe a screenshot would help.

To explain this sentence. I meant when watching a continuous recording with motion detection on from under the test and tune menu, if I see the orange frame around the entire clip viewer I would have thought I would find either a canceled alert or a confirmed alert clip corresponding so a few seconds before the time the orange frame comes in when I run the clip through test and tune motion detection. I dont.

I thought orange border here meant a trigger had occurred. But I don't see an a canceled or a confirmed alert clip is generated and there is no corresponding jpg in the alert folder.


My general question....if you see the orange frame when running a clip through test and tune motion detection, should I not expect to see either a canceled or a confirmed alert clip?

On a screenshot given this is general question not sure what I would show.
User avatar
YrbkMgr
Posts: 587
Joined: Sun Nov 24, 2019 12:56 am
Location: Chicagoland

Re: Test and Tune Motion and Deep Stack Review Orange Frame

Post by YrbkMgr »

iannecj wrote: Tue Aug 03, 2021 1:49 am To explain this sentence. I meant when watching a continuous recording with motion detection on from under the test and tune menu, if I see the orange frame around the entire clip viewer I would have thought I would find either a canceled alert or a confirmed alert clip corresponding so a few seconds before the time the orange frame comes in when I run the clip through test and tune motion detection. I dont.

I thought orange border here meant a trigger had occurred. But I don't see an a canceled or a confirmed alert clip is generated and there is no corresponding jpg in the alert folder.


My general question....if you see the orange frame when running a clip through test and tune motion detection, should I not expect to see either a canceled or a confirmed alert clip?

On a screenshot given this is general question not sure what I would show.
I think I see what you're getting at now. If I understand you right, I've experienced the same behavior and I can't figure out why. But just this evening, I noticed something... and it may be pure lunacy... but I'm trying to noodle it out. Here's what I noticed:

Reviewing Canceled Alert clips, I have the BI Status Window open with the DeepStack tab selected. When I double click on the clip in the clips pane, the Status window shows what DS detected at that time - based on what is in the clips.dat files that it creates when recording motion. In my case no car was identified, so no alert. But if I have that same status window open, and Analyze with DS from Test and Tune... when I double click on the clip again, the status windows shows that DS detects the car.

I repeated that process on all of my cancelled alerts where there were people or cars. In every circumstance, the Test and Tune / Analyze with DeepStack option identified the people or car, while the status window (from the .dat file) did not.

Since my observation is that DS analysis results are different when the .dat file is created vs real time analysis of the clip, I'm wondering if DS is analyzing the same way it does from the console. Can you replicate that?
  • "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
  • If you know what your after, you'll recognize it when you see it.
jrowberg
Posts: 2
Joined: Tue Aug 10, 2021 1:17 am

Re: Test and Tune Motion and Deep Stack Review Orange Frame

Post by jrowberg »

I have a similar (if not identical) problem on my end. I'm running the latest version at the moment (v5.4.9.17). DeepStack does amazingly well when working with the test+tuning analysis. This video shows exactly what the tuning output is using a clip directly from a cancelled alert with the "nothing found" classification:

https://www.youtube.com/watch?v=WJc7GyJ2T80

Ignore the forest sounds in the background. It's a white noise app. I get this kind of BI+DS behavior all the time, and I wish I knew why. DS clearly knows what is going on, but it almost never passes it on to BI correctly. And yet, the DS status window shows all kinds of success on this same clip:

Image

In global BI Settings, DS is configured with "Default object detection" enabled, mode "Medium", with timeout to 15 seconds. Facial recognition is off. I have tried "High" as well with no change in output. The camera has motion detection on, and I have a single zone configured to eliminate the street and include only the driveway as shown here:

Image

The native BI motion detection is configured to be extremely sensitive (min object size = 200, min contrast = 25, min duration = 0.5, "Simple" algorithm, High definition on), so that just about anything is passed on to DS for analysis. The camera's AI configuration is shown here:

Image

On very rare occasions, I do get alerts on this camera, but it misses tons of obvious instances like this. However, it seems that if a car goes DOWN the driveway instead of UP it, the probability of detection is better...but that seems ridiculous. Should it matter if the vehicle starts on the edge of the view inside the zone, as opposed to in the middle of the view and enters the zone from the top of the painted area? It shouldn't, but maybe it does. When I first set up DeepStack a few weeks ago, I turned off zones to see what would happen, but I way more alerts than I wanted because of the cars going by on the street out front. I figure a combination of DS and zones should be perfect, but alas...it isn't. Yet.

I just noticed that v5.4.9.18 is out today, including "Alert adjustments and fixes" and "DeepStack status window updates." I'll try the update right away to see if it fixes anything, but the last half dozen had the same changelog, so I'm not hopeful for a magic fix yet.
User avatar
YrbkMgr
Posts: 587
Joined: Sun Nov 24, 2019 12:56 am
Location: Chicagoland

Re: Test and Tune Motion and Deep Stack Review Orange Frame

Post by YrbkMgr »

I'm still researching but here's what I've concluded so far.

1. The analysis provided via the .DAT file is different than when using Test and Tune/Analyze with DeepStack

2. The documentation available to understand the DeepStack Tab in the Status Window is either sorely lacking or written for an audience of superior competency to mine. I have been unable to find how that Status can be used to understand what's happening.

3. The reason for the difference in object identification has been suggested to be some level of CPU choking - this is Tribal Knowledge from anecdotal observations from many users. A few have reported that using the GPU version of DS makes all of this go away. I cannot use the GPU version so I can't test it.

4. All alerts are handed off to DS but DS does not respect Hotspots (Zone H). Specifically, if motion occurs in a hotspot, DS cancels the alert unless it identifies an object listed in the AI section of the camera's Trigger Tab. I have gotten the impression that many users don't use hotspots and downplay the use of Zone H in general. Regardless, in BI, motion in a HS bypasses all conditions and fires an alert - it's an "all bets are off, trigger and alert". Deepstack will not behave the same way. Alerts are only confirmed when DS identifies an object. In my view, this is a problem - not THE problem, but A problem.

From what I've found, the most comprehensive way to view DS analysis is to have the Status Window open with the DS Tab selected - it should be blank. When you double click on the Clip, the DS window populates with the analysis done at the time of the event from data in the DAT file. If you then enable DS Analysis via Test and Tune, double clicking on the same Clip shows what DS finds in the BVR. As you've discovered, they are different.

I'm trying to get to the bottom of why they're different, and which one provides information that is of most value in fine tuning DS or camera settings. I'm going to write up my observations in a separate post, but I'm still collecting data.

So in summary... join the club of those confused by DS integration into BI.
  • "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
  • If you know what your after, you'll recognize it when you see it.
jrowberg
Posts: 2
Joined: Tue Aug 10, 2021 1:17 am

Re: Test and Tune Motion and Deep Stack Review Orange Frame

Post by jrowberg »

FWIW, my continued testing has not improved things. I've been running the .18 patch release since it was available (Aug. 9, 2021), but the behavior is the same. I've tried unchecking "Use main stream if available" to no effect, and I even tried disabling the substream entirely so that DS has only one stream to work with, again to no effect. I even tried increasing the global timeout (s) setting for DS to 60 instead of 15 seconds, but the behavior is the same. Manual review of canceled alerts continues to show perfect recognition of objects with high confidence from DeepStack, and yet the alert is still canceled.

I hope this can be fixed soon, or someone figures out where my configuration is broken.
User avatar
YrbkMgr
Posts: 587
Joined: Sun Nov 24, 2019 12:56 am
Location: Chicagoland

Re: Test and Tune Motion and Deep Stack Review Orange Frame

Post by YrbkMgr »

jrowberg wrote: Fri Aug 20, 2021 6:41 pm or someone figures out where my configuration is broken.
This is the area I struggle with as well. They need a white paper of sorts on: "Using DeepStack Status Window to Troubleshoot Falsely Canceled Alerts", or something like that.

There should be a general procedure to show folks where they look in settings to fine tune it. Camera firmware? BI Trigger settings? AI images/unit time? Something... As it is, for the average person, it's a "guess and check" exercise. I've been doing that since I upgraded - it's maddeningly frustrating.
  • "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
  • If you know what your after, you'll recognize it when you see it.
Fitz2380
Posts: 8
Joined: Thu Sep 24, 2020 10:59 pm

Re: Test and Tune Motion and Deep Stack Review Orange Frame

Post by Fitz2380 »

I was curious if YrbkMgr had been able to solve the original issue and if there was a follow up post related to this. I too have similar issues as originally noted with the FedEx truck and have been trying to determine why. I too have read all of the documentation and viewed the BI 'training' videos but still am at a loss sometimes when DS does not detect a vehicle.
User avatar
YrbkMgr
Posts: 587
Joined: Sun Nov 24, 2019 12:56 am
Location: Chicagoland

Re: Test and Tune Motion and Deep Stack Review Orange Frame

Post by YrbkMgr »

Fitz2380 wrote: Tue Jan 11, 2022 2:37 pm I was curious if YrbkMgr had been able to solve the original issue and if there was a follow up post related to this. I too have similar issues as originally noted with the FedEx truck and have been trying to determine why. I too have read all of the documentation and viewed the BI 'training' videos but still am at a loss sometimes when DS does not detect a vehicle.
I have solved the issue. It is a matter of understanding how to analyze information. This thread should be helpful but the summary is below.

There are two ways to view a DeepStack analysis: Review the DAT or Analyze the BVR. Each deliver different information. Use testing and tuning analyze with DS to review recordings and see what DS identifies. If you are trying to recognize dogs, and DS is identifying the dog in the video as a horse, you may want to include "horse" in the AI "to confirm" box.

Reviewing the DAT tells you what happened. See DeepStack tab about halfway down the page. That info will help you understand what the DS status window is telling you.

Give a gander to those, and I will help in any way I can.
  • "Whenever I take something apart to fix it and put it back together again, I end up with like six really important looking pieces left over" -Tim Allen
  • If you know what your after, you'll recognize it when you see it.
Post Reply