Add keypoint dataset class

[jveitchmichaelis]

Description

This PR adds a keypoint dataset that returns either (x, y) coordinates or a 2D density mask as target. The dict entry labels is consistent with transformers semantic segmentation models. This should cover the majority of the models that we might want to train. The structure is essentially the same as BoxDataset.

The diff is better than it looks. I've pulled out some duplicate logic for the box dataset and made base class which will be used for polygons as well. Most of the additions are in the test suite.

Models now have a task attribute which is used to load the correct dataset type (e.g. all current models have task = box). I think this is cleaner than having the user specify this via config, since it should be obvious by the choice of architecture.

Training is not supported, but I'd like to add that separately to reduce review load.

Related Issue(s)

#809
Supersedes #1182 as features from there are migrated.

AI-Assisted Development

Some suggest tests and refactoring.

• I used AI tools (e.g., GitHub Copilot, ChatGPT, etc.) in developing this PR
• I understand all the code I'm submitting
• I have reviewed and validated all AI-generated code

AI tools used (if applicable):

Claude Code, Gemini

[codecov]

Codecov Report

❌ Patch coverage is 87.85714% with 17 lines in your changes missing coverage. Please review.
✅ Project coverage is 86.87%. Comparing base (884502e) to head (18fe41b).
⚠️ Report is 8 commits behind head on main.

Files with missing lines	Patch %	Lines
src/deepforest/datasets/training.py	88.97%	14 Missing ⚠️
src/deepforest/main.py	62.50%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1311      +/-   ##
==========================================
- Coverage   87.35%   86.87%   -0.49%     
==========================================
  Files          24       24              
  Lines        2981     3176     +195     
==========================================
+ Hits         2604     2759     +155     
- Misses        377      417      +40     

Flag	Coverage Δ
unittests	86.87% <87.85%> (-0.49%)	⬇️

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]

Coverage looks OK. Some edge cases in the unit tests which we could add later, and main coverage can be ignored for now as we don't have models in yet.

[jveitchmichaelis]

Pausing review for a moment while I check some naming consistencies with utilities.

[bw4sz]

@jveitchmichaelis can you either take me off request, or take off draft, trying to organize the issue queue.

[jveitchmichaelis]

Minor change here - the density map is normalized to sum to the object count for each class.

[jveitchmichaelis]

@bw4sz this is good for review now, ignoring the pre-commit.ci failure.

[bw4sz]

I'm happy with the /src. I always am unsure about the amount of tests. For example, do we need flip tests? I think it's fine and they are pretty cheap to run. I'm going to approve, but something for others to weigh in on as we think about the size of the codebase, which I sense is going to swell a lot over the next year.

[bw4sz]

Happy to merge once the pre-commit is solved.

[jveitchmichaelis]

Agree. I added the flip test to make sure that we have coverage for augmentations, since we haven't really tested them with points. Dataset tests are at least very fast, as you say.

[bw4sz]

Please rebase main and that should be fixed.

[jveitchmichaelis]

Good to go when the CI passes

[bw4sz]

The #1343 made a tiny conflict, that I think you can handle right in GitHub. Then I will merge.

[jveitchmichaelis]

Rebased

[jveitchmichaelis]

All good, first rebase had a small conflict