IP Cam Talk Custom Community DeepStack Model

[sebastiantombs]

I've been using a single zone and dropped zone crossing a while ago. There were too many missed motions with zone crossing, especially with Edge Vector but even with Simple. DeepStack does a god job of filtering out the falses for me using just a single zone. I've never relied on Object Detection and have always used motion zones. It does take some tuning to get size and contrast set correctly for all lighting conditions, but again, DeepStack filters out the "false" trigger quite well.

 

[Lockwood]

I saw the same video and that's what I followed when setting up Deepstack. But I am also interested in what others are doing. I'm using the CPU and seeing decent response times with those. Haven't tried the GPU.

Cool, thanks for the reply. The fact that I'm 4 hours away from my camera makes it very hard to accurately set up Object Detection and motion triggering in general. So long as the GPU doesn't get burned out I'm fine with all the BI triggers being analyzed by Deepstack.

 

[Lockwood]

I've been using a single zone and dropped zone crossing a while ago. There were too many missed motions with zone crossing, especially with Edge Vector but even with Simple. DeepStack does a god job of filtering out the falses for me using just a single zone. I've never relied on Object Detection and have always used motion zones. It does take some tuning to get size and contrast set correctly for all lighting conditions, but again, DeepStack filters out the "false" trigger quite well.

Thanks for the response. I'll just BI be trigger happy and let DS do its job.

 

[Sybertiger]

Do you guys still keep these type of settings and allow Deepstack to analyze pictures based off a trigger from those settings or do you turn off Object Detection and increase the sensitivity of Min. Object Size and Min. Contrast and give Deepstack a better shot at detecting something based on those settings?

Not sure if I read that right but....



Maybe this this is old and obsolete or only applies if relying on the BI motion sensor only (not camera's internal alerts built into the camera's AI).

 

[sebastiantombs]

I leave Object Detection "ON" for major scene changes, lighting primarily, but use motion detection for the meat and potatoes end of it.

 

[Lockwood]

I had Object detection on to enable my Zones which I'm doing away with just like sebastiantombs did. I'm wondering if I need to just leave Object Detection on and turn off "Object crosses zones" instead.


A couple of more question regarding Cuda and cuDNN
1. I downloaded and will install Cuda 11.6. I'm assuming I need to download cudnn-windows-x86_64-8.4.1.50_cuda11.6-archive.zip? The Deepstack website says to use Cuda 11.3.1. Does Cuda 11.6 not work?
2. Those that have it all working. what versions are you using? Looks like in this thread, Help with deepstack gpu for windows, page 4, post #66 the user is using Cuda 11.3.1 and cudnn-windows-x86_64-8.4.0.27_cuda11.6-archive.zip. Maybe 8.4.1 wasn't out yet?
3. The current Windows install instructions for cuDNN found here, Installation Guide :: NVIDIA Deep Learning cuDNN Documentation, mention installing Zlib. The post in #2 above does not mention anything about that. Did you guys install it? If you did, where did you copy the .dll file and did you add that path to the PATH Environmental Variable?

Sorry for all the questions. I have limited time at my camera location and want to get this right the first time.

Thanks,
Chris

 

[carteriii]

Regarding the motion detection from one zone to another, I use Dahua cameras and have found that it's "IVS" line crossing is FAR more accurate and sensitive, so I use that as event triggers when I want that functionality. The BlueIris alert shows the trigger as "External" (rather than a zone), which is awesome. I thought (assumed?) Blue Iris then picks up the ONVIF trigger event and still calls DeepStack because the alerts will still often say "nothing found", which is why I believe/assume DeepStack is always called for external triggers, even if an object is not identified by BI.

Ironically, the "Artificial Intelligence" settings on the Trigger settings tab allows you to specify that DeepStack is called only for triggers in specific zones, but as far as I know it doesn't allow you to specify whether DeepStack is called or not based on an external trigger. I believe/assume that DeepStack is always called on an external trigger, no matter what. If that's wrong, someone please let me know.

 

[sebastiantombs]

The way I read the instructions on the DeepStack page the CUDnn version depends on the CUDA version which depends on the DeepStack version.

A trigger is a trigger whether it's from BI or an ONVIF event. BI "sees" the ONVIF trigger and treats it as a regular trigger from my experience.

 

[Lockwood]

@sebastiantombs - What about Zlib? Did you do anything with it in order to set up Deepstack?

 

[Sybertiger]

Sometimes I have those days with DeepStack which make me scratch my head and say WTF....LOL! Here's the analysis...at the very get-go DS at T+0 identifies the car and in another analysis it finds the car again but in the end it decided it didn't find anything. Why did it decide that nothing was found?

 

[sebastiantombs]

@Lockwood I simply followed the instructions to install found on the DeepStack page/guide. That doesn't mention zlib that I'm aware of.

 

[Sybertiger]

Anyone notice what a touch time DeepStack has with black colored vehicles? I have so may clips of nothing found when it's a black or very dark colored vehicle even in the daytime.

 

[wittaj]

Interesting, my DeepStack gets black colored vehicles day or night. Sounds like a contrast issue you need to work on within the camera

 

[Sybertiger]

Interesting, my DeepStack gets black colored vehicles day or night. Sounds like a contrast issue you need to work on within the camera

The camera tripwires and intrusion zones do a perfect job of sending the triggers to BI and DS. It truly is amazing what the AI in the cams are doing by capturing all color cars day and night. Cameras trigger every time. Not sure what kind of contrast DS needs.

 

[wittaj]

One would think it would see that, but I am wondering if the dark leaves are making DS think that is part of the vehicle and then misses it.

I guess the next question is why use the camera AI and DS? Is the only reason you are using DS is to get the orange indicator that it is a vehicle or is there another purpose and use case?

The only camera with AI that I use DS on is to get the whole vehicle in the alert image at night. Other than that, I simply let the camera AI do the triggering and alerting.

 

[sebastiantombs]

I've found that DS seems to need a high contrast image at night to reliably detect dark colored vehicles at night. It takes some tuning to get each camera "just right" due to differing fields of view and different lighting conditions. Back lit and front lit are tough while overall/overhead lit is fairly easy.

 

[bayernstern]

Folks, I've been running Blue Iris and DeepStack already for several weeks, using the custom models from @MikeLud1 (thanks again for the fantastic job!)... The specific models I am using are “combined”, “dark” and “license-plate”. Overall, I am extremely happy with how everything is working... These models translated into huge improvements for me in terms of enhanced triggering and false alert avoidance. That said, occasionally I do get false triggers... My question is: how one can train DeepStack to avoid thinking that, for example, my barbecue in the backyard is a person, etc. Thus far, I've been going to the AI Blue Iris Status window to assess the motion confirmation logic, and if incorrect, I manually cancel the alert and un-flag the specific trigger instance... That said, I am unsure if I am really training the model to become more accurate in any way, or if I am solely (manually) cancelling alerts and un-flagging without any positive impact whatsoever on my future DeepStack performance? Thanks in advance for any insights you may be able to provide! Very appreciative of this wonderful community.

 

[sebastiantombs]

Training a model involves working with the specific model, the model file itself, which is an involved process. Actions in BI have nothing at all to do with training a model.

 

[bayernstern]

Thank you @sebastiantombs for the swift reply... Bummer, I was afraid that was indeed the case...

Once again, immensely appreciative of all the insights of this fantastic community!

Unquestionably, I would not be where I am in terms of Blue Iris (including hardware and cameras) and DeepStack satisfaction without the wisdom of so many on here.

 

[loopy123]

Any ideas why I am not able to detect raccoons? Using the latest animal and combined models from first post.
See below, 5 raccoons mulled about my patio and never got a trigger. Note - my dog triggers this camera consistently.