Add validation checks for bounding boxes

[ethanwhite]

Checks to see if bounding boxes occur outside of image boundaries and clearly communicates to user when they do.

This is a version of #1015 that accounts for changes in the codebase over the last year. This was good work by @reddykkeerthi that I didn't want to see get lost.

Closes #1014

[ethanwhite]

@jveitchmichaelis I justed noticed that you have a similar commit somewhere from the history in #1015. We can close this if you were planning to submit something similar. Sorry I didn't catch that earlier.

[codecov]

Codecov Report

❌ Patch coverage is 90.00000% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 87.50%. Comparing base (c10846a) to head (c8a48c2).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
src/deepforest/datasets/training.py	90.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1211      +/-   ##
==========================================
+ Coverage   87.47%   87.50%   +0.02%     
==========================================
  Files          20       20              
  Lines        2586     2616      +30     
==========================================
+ Hits         2262     2289      +27     
- Misses        324      327       +3     

Flag	Coverage Δ
unittests	87.50% <90.00%> (+0.02%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[jveitchmichaelis]

I guess that's in the dino branch? Struggling to see where that commit belongs - it looks like I added it when I was testing the pretrain data and then squashed. So we can keep this one if it didn't make it into main.

[jveitchmichaelis]

The coverage checker is requesting failure tests for image path and geometry, might not be a bad idea to add those.

test_validate_dataset_missing_image and test_validate_dataset_bad_geometry etc.

[ethanwhite]

The coverage checker is requesting failure tests for image path and geometry, might not be a bad idea to add those.

Added a test for the missing image. Couldn't figure out how to trigger a bad geometry. You're welcome to PR one into this branch or point me in the right direction.

[jveitchmichaelis]

Yeah it's not clear to me how you'd fail that. Unless it's possible to store a bad box in WKT format and Shapely blindly loads the bounds? But otherwise I think the only way to trigger that error would be to have more (or less) than 4 bounding coordinates. Or if the shape doesn't have bounds defined - like a point?

[ethanwhite]

Yeah it's not clear to me how you'd fail that. Unless it's possible to store a bad box in WKT format and Shapely blindly loads the bounds? But otherwise I think the only way to trigger that error would be to have more (or less) than 4 bounding coordinates. Or if the shape doesn't have bounds defined - like a point?

Should I just go ahead and pull that try/except block then?

[jveitchmichaelis]

Apparently even a shapely.Point has "real" bounds: https://shapely.readthedocs.io/en/stable/reference/shapely.Point.html

Two options? If the check is to verify that it's a box, and we want to be strict, we should test for it geometrically. This seems like a neat solution. I would use envelope instead of allowing rotation.

https://stackoverflow.com/questions/62467829/python-check-if-shapely-polygon-is-a-rectangle

Alternatively we could take the bounds and assert non-zero area which would allow us to load polygon data (degraded as boxes), but not points.

Adding a check for zero (or very small) area is also probably good.