More cookbooks for quick start

[bw4sz]

User asks

we were wondering if you have any sample scripts that you share that e.g., load images, load annotations, train canopy delineation/extent model on image subset, run model on the entire image, assess accuracy of canopy delineation, attribute tree species based on field data collection, and assess accuracy of tree species. We can figure this out ourselves (possibly with some additional help from you), but I just thought I’d ask if you have a sample script for these tasks.

My sense is that all of this already exists within the docs, but the user started to look for it, couldn't find what they wanted and email me. @henrykironde this is part of the broader effort to improve navigation within the docs.
For each, let's point to location we would expect the user to find them in docs and figure out how to name and direct them there more cleanly.

• load images
• load annotations
• train model on image subset (i'm not sure exactly what subset means here, crop?)
• run model on entire image
• assess accuracy of model
• Multi-class point data to boxes
• Train Multi-class Model and Validate.

We could approach this as one long use case for a particular project in the example/ dir with jupyter notebooks, or we can make them tiny gists in a cookbook section. @ethanwhite thoughts?
Let's use this issue to keep track of the different places

[ethanwhite]

A few thoughts:

My sense is that all of this already exists within the docs, but the user started to look for it, couldn't find what they wanted and email me.

We need to work on offloading direct emails into the issue queue. My general approach to that is to have a nice, pre-prepared, email that basically says "Thanks for your interested in DeepForest. Here are the appropriate ways to contact the team. Please post this to GitHub as a ... and someone will get back to you as soon as they can." If we have users that can't use GitHub for some reason (e.g., country level domain blocks) then let's chat about that.

@henrykironde this is part of the broader effort to improve navigation within the docs.
For each, let's point to location we would expect the user to find them in docs and figure out how to name and direct them there more cleanly.

I agree that we have some work to do here and that it is one of our top priorities.

We could approach this as one long use case for a particular project in the example/ dir with jupyter notebooks, or we can make them tiny gists in a cookbook section. @ethanwhite thoughts?

Since this feels like this is the second most common use case beyond just using the models for zero-shot I think there could be some value in a larger step-by-step examples for: 1) first 5 steps focused on detection; and 2) training and validating a multi-class model. I think point-to-box conversion should just be a standard docs item.

[colbyrand]

Perhaps this is already somewhere in the docs, and I can't seem to find it, but is is there an example script that you could share that touches on the following points:

• loads images and annotations for both the training and validation areas
• runs the default model on an image and explains what model parameters you can change and where to change those in context of forest conditions (e.g., open v closed canopy)
• trains the canopy delineation model to a subset of the image based on the annotations you supply and validates model with a second set of annotations
• Runs the trained model on the entire image and exports the predictions and model accuracy assessment
• Use field validated training and test data to predict tree species and assess accuracy of tree species prediction.

Thanks!

[bw4sz]

Good example from supervision

https://roboflow.github.io/cheatsheet-supervision/