I posted this as a comment to another thread, but I wanted to discuss it more directly.
It’s surprising that DeepStack detects Elephants and Giraffe but not more common animals like deer/elk, rabbits, fox, etc
We have 21 cameras on our rural property and our land is swarmed with deer, jack rabbits, red fox, bobcats, etc, so using DeepStack is just not an option for us unless we have a herd of stampeding Elephants running across the land.
I’m sure it’s probably difficult to detect all specific animals, but it would be at least nice to have a fuzzy detection that something is more than likely an animal. I would like to avoid vehicles and just detect people and animals, but as it stands, that’s not going to be possible, so I’ll have to stick with Blue Iris’ built in detection and hope DeepStack gets more fine grained detection in future.
It’s not difficult, already done but can’t be used with deepstack so far: MegaDetector
Please work on it with priority: https://forum.deepstack.cc/t/import-third-party-models/555
Have a look up to see where the developers are and then you may understand.
I would also like to see more common animals added.
Deer, squirrel, raccoon, opossum would be my top choices. They are frequently tagged as other animals, but if a miss one deer my garden could get wiped out. I made a deer model that works when tested, but isn’t doing much in the app.
@Glen_Cove, @bmorgan99 , @pentium90, @Tinbum
Please check this response.
Turns out I had ai tool set incorrectly and my deer model is working now. As to Moses’ reply, I’m sure we are just hoping someone else will do the modeling as it’s a bit tedious and a few of the included choices might not see much action.
Can you share this model?
The latest version tagged a spider web and moth as deer, so I’m gonna drop back a version and test. I will share it when I get something usable.
EDIT: This one is working so far. I’ll trade ya for a squirrel.
Still had problems with spider webs, so I’m trying model yolov5l. Doesn’t seem much slower to me.
It’s more accurate with my test files, but won’t know for sure until RL at night.
@bmorgan99 The most accurate model is the yolov5x . If you absolutely prioritize accuracy over speed, I will suggest you opt for this version of the model when training.
Are you using Blue Iris? The custom model seems to the right path, but the newest version of BI does not call the custom object API.
I’m using AI Tool with Blue Iris.
I would LOVE to see a model repository created where folks could download needed models. Like I’d love to see some more common animals as well – especially predatory – so I can be alerted when any of these enter any camera view on watch. In addition I would love to see some models of delivery service vehicles (UPS, Amazon Prime, FedEx, etc.) so custom announcements could be used. I realize I could make my own models, but after reading all the docs I don’t feel comfortable enough even attempting it. Many people would benefit from a model repository though - especially those like me that utilize and grasp a great deal of the technology, but remain out-of-range for things like custom model creation. If enough people were willing, I’d even host the platform.
Curious - what items are included in this model? Is it just “deer”? Is there a way to see what objects are included?
So which deer .PT is the best one that you are currently using. I have constant issues with deer crop destruction and currently I’m relying on it being picked up as a sheep or bird.
I’m using .9 currently but I’m not sure it’s much better. I haven’t have any plant damage all summer, but I’ve noticed it takes a good while to get a confirmed hit. The deer are sometimes in the garden.
When I run tests using the the dev kit and python, all my images are nailed quickly. My problem could well be with the Blue Iris config. Have you tweaked that at all?
Here are results from the first 2 semi-working months. I’ve since gotten a better camera and tweaked things a bit.
@bmorgan99 I have been working on custom models for DeepStack to use with Blue Iris. Can you share the the images and labels files you used to create your custom models so I can add them to the models I am working on. Below is a link to my GitHub repository for my models.