IP Cam Talk Custom Community DeepStack Model

[T_Tronix]

I created a pt for squirrels but my Deepstack never triggers. How do you combine images into one pt? What app did you use?

 

[MikeLud1]

I created a pt for squirrels but my Deepstack never triggers. How do you combine images into one pt? What app did you use?

I used FiftyOne to download the COCO dataset which is what DeepStack uses for their object detection. With FiftyOne you can filter out the images & objects you what them add the new image for the objects you want to add. After you have the new dataset made then you use the new dataset to train the new model.

Link COCO Integration — FiftyOne 0.14.3 documentation

 

[allgrind]

Hi, everyone,

I'm a little unsure - have I set everything correctly? Since I use the CPU version of Deepstack, only the Combined version should be used.

Should I therefore disable the setting: Default object detection?


Garden - camera settings



Many greetings
Jan

 

[sebastiantombs]

Looks good, but uncheck "Use mainstream if available". By default the image is down sized, resolution wise, for DS to work with it.

You might reduce the "analyze each" to 500 or even 250ms depending on the loading of your CPU. You may also want to increase the number of images to be analyzed. Both the time and number of images are dependent of the scene, speed of the objects and lighting in the scene.

 

[LJR_2019]

Any idea why I keep getting 'nothing found' when I use the combined model pack? I'm running deepstack on a linux server.

I start the server with:
sudo screen -d -m -S ds nice -10 docker run --gpus all -v /home/leo/my-models:/modelstore/detection -e THREADCOUNT=9 -v localstorage:/home/leo/deepstack -p 81:5000 deepquestai/deepstack:gpu-2022.01.1

 

[digger11]

You have "Save DeepStack analysis details" checked so you should be getting .dat files created in your alerts folder. When you drag and drop one of them onto the DeepStack tab of the BlueIris status window does it show that the combined model was used?

 

[Corvus85]

Am I correct in thinking that the 'dark' models are included in the 'combined' file? Or do I have to download another file just for night footage?

 

[MikeLud1]

Am I correct in thinking that the 'dark' models are included in the 'combined' file? Or do I have to download another file just for night footage?

I have not added the dark images yet to the customer models. I have been working on the DeepStack ALPR. If the combined is identify the items you need then no need to add the dark model.
I will start work on updating the custom models sometime next month.

 

[nz1984]

Any idea why I keep getting 'nothing found' when I use the combined model pack? I'm running deepstack on a linux server.

Yeah, I'm getting possibly the same thing trying to use it on a linux server too (link)
My speculation is that Blue Iris isn't correctly forming the URL when trying to access a custom model on a remote server (as opposed to how it forms the URL for the local install approach).

 

[Corvus85]

I have not added the dark images yet to the customer models. I have been working on the DeepStack ALPR. If the combined is identify the items you need then no need to add the dark model.

No problem.
I ask because I've noticed it misses very obvious objects at night - such as cats which walk diagonally from one corner of the frame to the other.
Where can I find the latest version of the dark model?

I will start work on updating the custom models sometime next month.

Much appreciated!

 

[MikeLud1]

Where can I find the latest version of the dark model?

GitHub - OlafenwaMoses/DeepStack_ExDark: A DeepStack custom model for detecting common objects in dark/night images and videos.

A DeepStack custom model for detecting common objects in dark/night images and videos. - OlafenwaMoses/DeepStack_ExDark

github.com

 

[MikeLud1]

I ask because I've noticed it misses very obvious objects at night - such as cats which walk diagonally from one corner of the frame to the other.

Setting the DeepStack mode to High will help identify smaller objects.

 

[Corvus85]

Setting the DeepStack mode to High will help identify smaller objects.

Yeah, already done :-\
I'll see if the dark model improves things.

 

[Corvus85]

Have I got this configured correctly?
I want to use both the 'combined' model and the 'dark' model for this camera.



Previously, I had 'combined, objects:0' in the 'custom models' field, but I want to add the 'dark' model.

But now, I get no confirmed alerts at all, and just get 'DeepStack: 404' errors on my cancelled alerts.

 

[sebastiantombs]

I'd flip that around to "objects:0,combined" just for a laugh.

 

[Tinman]

Also try no spaces between models.

 

[Corvus85]

I'd flip that around to "objects:0,combined" just for a laugh.

Also try no spaces between models.

Did these - no luck. Still shows 'DeepStack:404' and goes into the cancelled alerts :-\

For more info: I've put the 'dark.pt' file in the same folder as the 'combined.pt'. Is this right?

Is there any other configuration I needed to have done? I don't remember doing anything else with DS when copying over the combined model.

 

[Corvus85]

Ok so after not even changing anything, I have alerts being confirmed, but it still says 'DeepStack:Timeout' underneath them.



How is that even possible? How can it timeout if it's already successfully detected a person in the frame?

 

[wittaj]

Take zebra out of the cancel to test it. Also only run one model to test.

When you have something in the cancel box and running two models, it will then run every image that you have asked it to, which in your case is 10 images, through all the models you asked. But it is taking too long to do all 10, so it times out.

 

[Vettester]

Ok so after not even changing anything, I have alerts being confirmed, but it still says 'DeepStack:Timeout' underneath them.

View attachment 116832
View attachment 116833
How is that even possible? How can it timeout if it's already successfully detected a person in the frame?

Try changing your mode back to medium and lowering the number of +real time images to 2. You may need to raise the analyze one each to 750ms as well.