Deepstack Missing Detection

[Lockwood]

Hey everyone,

I'm new to Deepstack and just got it up and running last weekend. I seem to get a lot of "Nothing Found" when there is in fact movement. Please take a look at the attachments. There is a deer in both pictures and that deer moves from left to right in one pic and right to left in the other. Both show as nothing found in BI, but if I analyze the alert in Deepstack, it finds the deer in its complete path. Although Deepstack sees the deer as a deer and a cow and a dog and a fox. That doesn't bother me, so long as it finds something and shows it in BI. The final analysis just shows Cow with a confidence level of 88% in one pic and shows deer in the other with a confidence level of 73%. There are a bunch of other analysis attempts that all show Nothing Found. The final result shown in the BI alert is nothing found.

The To confirm field contains person, vehicle, bird, cat, dog, bear, raccoon, rabbit, fox, skunk, squirrel and cow.
The To cancel field contains Horse since there aren't horses around and this should allow Deepstack to analyze the entire picture.

Do you guys see a setting I should change? Should I increase the confidence level? I'm using sub streams, should I try and have the main stream analyzed?
I'm running BI 5.5.7.11 and Deepstack for GPU version 9/1/2021

Thanks,
Chris

 

[sebastiantombs]

Two comments.

Contrast is always a factor. Maybe try tuning the cameras for a little more contrast during the day. At night contrast is very critical and it can take some time to get things just right.

Try removing "horse" for the "To Cancel" box and just leave it blank.

Remember that "testing and tuning" is not exactly the same as the detection used on live video. With that in mind use the file written by saving the analysis details to see what is going on rather than using "Tuning and Testing".

 

[Lockwood]

Thanks sebastiancombs, you've always been very helpful. I already have the contrast set to 25 for these cameras, but I can definitely lower it. I'll remove horse and see what happens. Other than that, do the settings look ok?

Also, I'm going to my Alerts folder and clicking on the .DAT file (I assume that's what you meant when saying use the file written by saving the analysis detail) and I get the attached error. The weird thing is I have the DVR format set for all my cameras. Any clue why BI won't play it? That is if I'm even doing it correctly.

Thanks,
Chris

 

[sebastiantombs]

To use the .dat file open the BI console and click on the graph icon with the up arrow/lightning icon on the top left. Select the DeepStack tab and drag the .dat file into the lower pane of that window.

Your other settings look fine. Increase your contrast a little bit at a time. Don't rush because it's easy to overdo the increase. That should help.

 

[Lockwood]

Thanks, appreciate the assistance.

 

[sebastiantombs]

Hope it helps. Let us know how it works out.

 

[Lockwood]

I made the suggested changes, we'll see what happens.

The way you describe to access the .dat file. That's the same as finding the alerts, holding down Ctrl and double-clicking, correct?

One more thing. What I don't understand is using the .dat file, see attachments in first post, it shows that a cow was found with 88% confidence and a deer was found with 73% confidence. If Deepstack found those and they were both in the To Confirm field and my confidence setting is set to 40%, shouldn't have Deepstack shown in BI that a cow was found and a deer was found? Why would it show nothing found rather than the animal it found? It almost seems as if Deepstack is ignoring the animal model and only going off the General model.

I'll continue to tinker and see how it reacts.

 

[sebastiantombs]

I usually open the DeepStack tab, then open file manager and navigate to the alerts directory. From there I locate the file I want based on the camera name and time, drag and drop. That may be the long way around the horn, but it shows the actual results of the detection. I believe holding the control key and double slicking actually runs the clip through "testing and tuning" which is not the same as what happens during actual detection in alerts.

 

[Lockwood]

Took me a while, but I did some more testing. I looked at the .dat file by opening Blue Iris, clicking on the graph symbol and dragging the file into the lower right window on the Deepstack tab. This shows the same exact information if I hold the Ctrl key down and double-click on the Alert. I tested with the Alert and then found its corresponding .dat file. Both ways yielded the same information line for line.

That being said, I still do not understand how Deepstack comes up with a final result that it shows on the Alert in BI. Looking at my examples attached above, Deepstack found some animals and they are in green as a positive identification with confidence levels well over 40% which is what I have my confidence level set to. You would think Deepstack would use that positive identification and show that something was found in the picture, such as the cow at 88% or the deer at 73%. Instead it chooses to say nothing found. This is causing me to miss Alerts.

Fast forward to today. As you can see from the attachments below there is a deer walking around. Brown on green which stands out so I don't think it is a contrast issue (contrast set to 25). Obviously Blue Iris detected the contrast difference as it triggered an Alert. Deepstack analyzed the alert and its final analysis is nothing found. The attachments below are the analysis for this Alert. I had to grab four different captures, but this is what you would see if you you scrolled from top to bottom in BI. I just named it top, middle A, middle B and bottom (as if scrolling from top to bottom). You can see at certain points, Deepstack finds a cow and at different confidence levels. Most of which are above my set confidence level of 40%. This is the part I don't understand. A cow is being found, but Deepstack has a red mark next to it like it wasn't found and gives a final analysis of nothing found.

Why does Deepstack find things (in green), but give a final anaylsis of nothing found. Why does Deepstack find things, but mark them in red and give a final analysis of nothing found? I'm just trying to understand how Deepstack decides on its final result to tell BI. Maybe if I understand that, I can understand my results.

I have begin analysis with motion-leading image checked. Do you think this may have an effect on the results? Should I uncheck it?

Thanks,
Chris

 

[Lockwood]

Here is an example I just grabbed from a different camera. The final result is Bird at a 71% confidence level in BI. As you can see, that was the highest positive identification in the analysis. So it got it right with this camera. This got me to thinking. The camera that Deepstack is having issues with is Reolink. Didn't Reolink have issues with the the iFrames being 1/2 rather than 1 to 1 or something like that? See the attachment with the .5 circled. I thought I read that this would mess up Deepstack. I have newer firmware that I believe fixes this issue and makes it 1 to 1. I'll install that and test again.

 

[OneCalledWell]

Here is an example I just grabbed from a different camera. The final result is Bird at a 71% confidence level in BI. As you can see, that was the highest positive identification in the analysis. So it got it right with this camera. This got me to thinking. The camera that Deepstack is having issues with is Reolink. Didn't Reolink have issues with the the iFrames being 1/2 rather than 1 to 1 or something like that? See the attachment with the .5 circled. I thought I read that this would mess up Deepstack. I have newer firmware that I believe fixes this issue and makes it 1 to 1. I'll install that and test again.

Just an fyi.. I am having the exact same issue with Duhua cameras. I have a total of 8 cameras, 6 of them them dahua and i will get a person aleart about 10% of the time from the dahua cameras. Also, Same as you, I run the video through the analyser and it says 80-100% confidence that its a person but I never recieve an alert. If you ever figure this out please let us know.

 

[Lockwood]

@OneCalledWell - The issue was happening mostly with my Reolink camera. I have read that the issue is Reolink cameras do not allow you to adjust the iFrame. It should be 1 to 1, but with Reolink cameras, it is 1 to .5. Using sub streams is even worse, it is 1 to .25. Deepstack needs it to be 1 to 1 or it may not detect objects. I wish I could find the thread, I was just reading it again yesterday. I have checked the box to have Deepstack analyze the main stream and I'm getting better results. Unfortunately the new firmware I updated the camera to still does not allow for iFrame adjustment. So .5 it is for now.

Having said that, I still experience some issues here and there with a deer or an animal walking across my lawn and Blue Iris detects it, creates an Alert and Deepstack analyzes the Alert. Deepstack says nothing found on the Alert in BI, but If I Iook at the .dat file it creates in BI I can clearly see Deepstack has found the deer (although it labels it as a dog or cow sometimes) with a confidence level greater than what I have my confidence level set to. It shows a green check, the animal it found and the confidence level, but gives a final result of nothing found. Although, I have to say, it seems to be getting better.

I just wish I knew how Deepstack worked. How does it determine the final result? Sometimes I see green checks and the final result is nothing found. Other times I see green checks and the final result is correct. Other times I see yellow clocks? with a confidence level, what does that mean?

I know you can set the iFrames in a Dahua. Make sure you set the I frame Interval to equal the Frame rate (FPS). Also, if using sub streams, do the same and make sure the Frame rate of the sub stream matches that of the main stream. See my Dahua settings below. This is the advice I have been seeing in the many threads I have been reading on BI and Deepstack.

Chris