A concept DeepStack YOLOV5 model for detection emergencies in my valley from my tower / sky cams. Mainly fires and Starflight fly overs.
See Convert and Merge example for details on its creation and 0.X versions.
Note training images ck*.jpg are from the Kaggle WildFire-Smoke-Dataset-Yolo dataset though the mappings are mostly altered to better align with my detection goals.
To use the model download the fire.pt
file to your C:\DeepStack\MyModels
folder. (Can be any folder but Blue Iris defaults to C:\DeepStack\MyModels)
-
Add a folder inside your DeepStack instance (If you run on Host) or mount it in Docker.
-
Run DeepStack with the command
--MODELSTORE-DETECTION "C:/DeepStack/MyModels"
See DeepStack utils for more info on setting up and testing your set up. -
To use RMRR detection call the end point
/v1/vision/custom/fire
with your picture and you get a response similar to this if it finds a known object:
{
'success': True,
'predictions': [{
'confidence': 0.93365675,
'label': 'fire',
'y_min': 279,
'x_min': 640,
'y_max': 340,
'x_max': 767
}]
}
The list of currently trained for objects in in fire_classes.txt
trainTest results are in train/trainTest.results.txt
debug.pics contains training run debug images highlighting where the model did not fully align with expected results, suggested map changes and such.
For more details on training see DeepStack training
For more help setting and or debugging a DeepStack setup see Quick Blue Iris with DeepStack debug
See runTrain.bat for an example of running training locally with my DeepStack Utils repo style setup.