IP Cam Talk Custom Community DeepStack Model

[MikeLud1]

thanks, i got V2.1 general installed now, seems to be working. Altough not as fast as i would like. Would you suggest using main stream or just sub stream data on high model (Nvidia P2000 deepstack gpu windows bi machine)

What detection times are you seeing.

 

[aadje93]

Allright, got it setup now on sub-high model. Are there any cuda/deepstack/driver tweaks i can do to further improve the performance? I dont see any gpu % hit when deepstack runs, so it seems like its more of a memory bandwith thingy instead of gpu load

 

[karlocikovic]

After few minutes of testing, I can confirm that this model does better than combined 2.0.

I have Reolinks B800 cams (streaming to BlueIris via Neolink), nVidia 1030 with 2GB ram, and i5 9600. GPU Deepstack, High mode, substreams from my "shitty" Reolinks ("Balanced" or externStream"). Never missed any motion with this combo. Look at times... amazing.

Thank you Mike!!!

 

[Charles_CO]

Quick question about DeepStack. I've been testing animal and combined models specifically to see if I can get some decent animal detection.
So far it's been useful most of the time when comes to deciding between animal and human. That said I've been identified as animal a few time.

I don't have luck with animals. When I have a deer in back yard I get confirmation on horse, cow, fox but deer is not even considered once. I guess our deer are unique. LOL I understand it very much depends on the setup and conditions. This was during the day and really nice day too.

Anyway, since I have pretty bad luck with this and I don't want to freak out every time when I have "hors" or "cow" in backyard... Is there any way within DeepStack to group animals and get confirmation on animal rather than on the name of the animal which is never correct in my case?

I really just need to be able to get confirmation on human and on animal. I can figure out what animal we are dealing with once I see the footage.

Any suggestion?

@MikeLud1 - How hard is it edit model and rename individual animals to animal if there is no way to group them sort of like function in BI > Mark as vehicle

It would be nice get notification on animal in general rather than wrongly identified animal.

Thanks Charles

 

[bqz]

Hi!

Mikelud, brilliant work! I'm using "general" and it has a very good performance with deepstack gpu.
May I ask you if you considered creating a custom model for "person" only?

Thank you for sharing your work!

 

[sebastiantombs]

@Charles_CO I don't think there is a way to edit a model once it's run through training. It would take another full training session with snapshots identified as the object(s) you want to be detected. I could be wrong about that, but seriously doubt that I am.

 

[Charles_CO]

@sebastiantombs Got it! Well, I have created another camera copy that will use animal model only and I will use is for actions that need to be taken with animal intruders. LOL
This way I can customize it the way I want and sent trigger to home automation system when "hopefully" animal is detected.
Flexibility of BI is amazing!

Hopefully AI will advance and gets better with recognizing objects. It would not be bad idea to somehow implement min and max size of the object in the frame. It would eliminate some false positives.

 

[MikeLud1]

Is there any way within DeepStack to group animals and get confirmation on animal rather than on the name of the animal which is never correct in my case?

I am thinking of making a new model with only three labels animal, person, and vehicle

 

[MikeLud1]

@Charles_CO I don't think there is a way to edit a model once it's run through training. It would take another full training session with snapshots identified as the object(s) you want to be detected. I could be wrong about that, but seriously doubt that I am.

Correct.

 

[Charles_CO]

I am thinking of making a new model with only three labels animal, person, and vehicle

I think that makes sense. Recognition is not perfect yet to reliably distinguish between different animals but if we can get it to the point where we can reliably separate animals, people and cars, that would make most users satisfied IMO. I think that is what most people probably look for anyway.
Also, it might possibly speed up things a little bit too.. Maybe.

 

[Futaba]

I am thinking of making a new model with only three labels animal, person, and vehicle

I have a RTX3090 on another computer running AMD 5950x CPU. If I can figure out how to run training, maybe I can help with training lower priority/less important models.

 

[MikeLud1]

I have a RTX3090 on another computer running AMD 5950x CPU. If I can figure out how to run training, maybe I can help with training lower priority/less important models.

If you want to give it a try, you need to install CUDA & cuDNN and follow the Local Setup in the below link. I am using a RTX3060 ti for local training.

GitHub - johnolafenwa/deepstack-trainer: Custom Object Detection Training for DeepStack

Custom Object Detection Training for DeepStack. Contribute to johnolafenwa/deepstack-trainer development by creating an account on GitHub.

github.com

 

[Futaba]

If you want to give it a try, you need to install CUDA & cuDNN and follow the Local Setup in the below link. I am using a RTX3060 ti for local training.

GitHub - johnolafenwa/deepstack-trainer: Custom Object Detection Training for DeepStack

Custom Object Detection Training for DeepStack. Contribute to johnolafenwa/deepstack-trainer development by creating an account on GitHub.

github.com

Do you do your training on Linux or on Windows in Power Shell?

 

[MikeLud1]

Do you do your training on Linux or on Windows in Power Shell?

In Windows with python

 

[aadje93]

I am thinking of making a new model with only three labels animal, person, and vehicle

if you would like to do this, or maybe give a link to the sample data so i could train my own set (P1000 + GTX 1080 ti) then i will happily make a "general light" with just person/vehicle/animal.

As i dont need to know if it was a truck, car or bus that loaded my stuff. Just a vehicle is enough for ALPR cam

Better to run 30-40 images through a razor fast deepstack than 3-4 as it takes to long.

 

[MikeLud1]

if you would like to do this, or maybe give a link to the sample data so i could train my own set (P1000 + GTX 1080 ti) then i will happily make a "general light" with just person/vehicle/animal.

As i dont need to know if it was a truck, car or bus that loaded my stuff. Just a vehicle is enough for ALPR cam

Better to run 30-40 images through a razor fast deepstack than 3-4 as it takes to long.

@aadje93 the original DeepStack object model was trained using the COCO Dataset. You can use FiftyOne to download and filter out the labels you do not want. Most of my custom model I used the COCO dataset

 

[aadje93]

@MikeLud1 thanks for the links, I will look into it

 

[mangocrysis]

sub stream data on high

When you say use "sub stream", do you mean unchecking "Use main stream if available" in trigger settings?

 

[MikeLud1]

When you say use "sub stream", do you mean unchecking "Use main stream if available" in trigger settings?

Yes

 

[MikeLud1]

I decided to upgrade my GPU that I use for training the custom modes. I found a good price on an open box Asus ROG-STRIX-RTX3090-O24G-GAMING at Micro Center (the GPU still had the plastic on it). Normally they go for $1,699.99 the open box price was $1444.96 with full warranty.
I fixed my local DeepStack Trainer now with this new GPU the training times are 4X faster then Google Colab Pro and 2X faster then my old RTX 3060 ti.
This weekend I will try to do some more model work.