Can we do anything from our side to improve DeepStack?

[ProTapper]

Hi guys,

QQ, we collect so much good data from our feeds. Is there something we can be doing to improve or help the DeepStack learning with sharing or submitting our data anywhere for improved model training? If so, what can we do?

Sorry if this is a dumb question. Thanks.

 

[aesterling]

Download updates as they become available and train a custom Deepstack model if you’re feeling adventurous.

 

[wittaj]

Yep, we can take our own pics and train our own model for our field of view to improve the accuracy at our locations, at least for those of you that can get the custom models to work LOL....

 

[ProTapper]

Haha I suppose that's obvious... I wasn't sure though if there was a crowd sourced community of their own improving the core models they ship with their install packages.

For the custom models... Does it require a powerful machine? How are you guys doing it?

 

[wittaj]

There is the DeepStack forum where folks into that sort of thing are doing those types of mods.

DeepStack Forum

Learn, share and discuss on the DeepStack Platform

forum.deepstack.cc

Training the models is "just" taking your photos and delineating what you want it to recognize. I haven't done that yet, but someone here did for plates and he said it took a couple hours to delineate around 100 photos.

Training Your Custom Model

docs.deepstack.cc

The main processing is in the live implementation of it. Obviously the more cameras, MP and cameras you are using for DeepStack, the more powerful the machine needed LOL.

 

[ProTapper]

There is the DeepStack forum where folks into that sort of thing are doing those types of mods.

DeepStack Forum

Learn, share and discuss on the DeepStack Platform

forum.deepstack.cc

Training the models is "just" taking your photos and delineating what you want it to recognize. I haven't done that yet, but someone here did for plates and he said it took a couple hours to delineate around 100 photos.

Training Your Custom Model

docs.deepstack.cc

The main processing is in the live implementation of it. Obviously the more cameras, MP and cameras you are using for DeepStack, the more powerful the machine needed LOL.

You're everywhere with your wisdom, AJ! I'll let the big boys then keep training haha! When a bird or a spotlight was identified as a person for me... I wondered if there was a way to give back? Haha!

 

[Skispcs]

You can make a custom model.
About 350 pictures and you need to put a box around the object and label it in each one using the LabelIMG software.

I have been experimenting with one to identify wildlife at my house.

 

[ProTapper]

You can make a custom model.
About 350 pictures and you need to put a box around the object and label it in each one using the LabelIMG software.

I have been experimenting with one to identify wildlife at my house.

Are custom models now working with BI in a stable way?

By the way, do you have any tips on catching critters like a possum, skunk etc work DeepStack? Is the a keyword for them?

 

[Coldair]

You can make a custom model.
About 350 pictures and you need to put a box around the object and label it in each one using the LabelIMG software.

I have been experimenting with one to identify wildlife at my house.

that sounds interesting

 

[Dave Lonsdale]

I am still way behind most other forum members when it comes to getting what I want out of the integration of DeepStack with Blue Iris. Perhaps different to most, I’m still at the basic first step - reliably confirming triggers. (Right now I don’t send alerts). Step two will be adding custom models/wildlife not currently in Deepstack’s database and perhaps cancelling known objects (like me), but despite having been fortunate in that fenderman reviewed my settings, I’m still in the situation where many objects already in Deepstack’s default database are cancelled.

My guess is a principal reason for many false cancellations is when an object is only in view or valid for a short time. Is BI is only able to send key frames to DeepStack? If so, having already been triggered using a perfect complete image, doesn’t BI have to wait up to another second before an image is sent to DeepStack for confirmation? (key frame interval of one second assumed). Sometimes this is too late and BI’s valid trigger is cancelled. I did ask this question previously (5.4.6 - May 17, 2021 #24) but was unanswered:-

5.4.6 - May 17, 2021

There's a bug where alerts don't recognize faces starting with capital letters. So "face1" and "face*" will work, but not "Face1" nor "Face*". Also, another bug where renaming faces to change capitalization doesn't actually register in the system as a change, so e.g. renaming "face1" to "Face1"...

ipcamtalk.com

If this is true, wouldn’t it be better if BI sent the image it already has causing the trigger to DeepStack?