Deep Learning Tree Detection model came out with results that seem "too sensitive"

[AndrewWallick]

I'm trying out the Deep Learning Tree Detection model for the first time, using mostly the default settings (I did change the Cell Size to 0.25). But I'm very new to Image Analyst and imagery analysis in general, so I don't understand what most of these settings mean. It looks like the model was too "sensitive" when detecting trees, but also most of the boxes that are correct look too small - like they don't contain the entire tree.

How should I try adjusting the settings for my next test of the model?

Also, if there's any blogs or other documentation that answers these questions, let me know. I read some things but didn't find much so far.

[PavanYadav]

@AndrewWallick maybe try with:

1. a few different thresholds (I see you're using 0.1; meaning if the model is 10% confidant it will return a result)

2. different cell sizes. 

3. test_time_augmentation = True

[ShivaniPathak]

Hi @AndrewWallick,

Could you please share a screenshot showing the tree labels you used when exporting the training data? Additionally, which model did you train?

You might consider trying the DETReg model, a few-shot learning model that excels at learning from a limited number of samples compared to traditional object detection models. DETReg (Detection with Transformers for End-to-End Region Queries) leverages the power of transformers and a self-supervised pretraining strategy, making it highly effective for scenarios where annotated data is scarce. This approach enables it to perform well even with minimal training data while maintaining high accuracy in object detection tasks.

For a practical example, you can refer to this sample notebook, which demonstrates how to train a DETReg model for palm tree detection: https://developers.arcgis.com/python/latest/samples/detecting-palm-trees-using-deep-learning/.