[niksirbi]

Draft skeleton for running point trackers from poseinterface, with each tracker in its own isolated uv env. Signatures and types only; nothing is implemented yet.

This was the result of an initial brainstorming session involving myself, @lochhh, @sfmig and @claude.

What's here

• poseinterface/trackers.py: PointTracker ABC (no torch). Subclasses implement track(video, queries) -> (tracks, visibility); the shared run() reads PoseMark start labels and writes predictions as PoseMark clip labels.
• envs/cotracker3/: standalone uv project (path dependency on poseinterface, not a workspace member, so each tool keeps its own lockfile) with a CoTracker3 skeleton and CLI.
• Tests: base class in tests/test_unit/test_trackers.py (main env); tool tests in envs/cotracker3/tests/, run via uv run --project envs/cotracker3 pytest envs/cotracker3/tests.
• Root pytest restricted to tests/; envs/ excluded from the sdist.

See HANDOFF.md for design decisions, next steps, and maintainability notes (to be removed before merge).

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 75.00000% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 93.45%. Comparing base (0df45e1) to head (26ad153).

Files with missing lines	Patch %	Lines
poseinterface/trackers.py	75.00%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #86      +/-   ##
==========================================
- Coverage   93.99%   93.45%   -0.54%     
==========================================
  Files           5        6       +1     
  Lines         416      428      +12     
==========================================
+ Hits          391      400       +9     
- Misses         25       28       +3     

☔ View full report in Codecov by Harness.
📢 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.

[lochhh]

Alternative design to #85

Each tool runs in its own isolated uv environment, e.g.

poseinterface/            # main env: import, export, convert, compare
envs/                     # tool-specific envs
  dlc/  pyproject.toml  uv.lock  train.py  infer.py
  lp/   pyproject.toml  uv.lock  train.py  infer.py
  <tool>/ ...

• One env per tool: no cross-tool dependency conflicts, and adding a tool doesn't touch the others.
• Each tool has thin wrapper scripts (train.py, infer.py) that call the tool's own API/CLI. Wrappers share a core set of arguments to allow automating looping over tools.

Workflow

1. Import labels to PoseMark (main env): Convert existing labels (DLC, LP, SLEAP, ...) to PoseMark with annotations_to_posemark and related functions, define the Train/Test splits.
2. Export Train split to tool formats (main env): Convert the PoseMark Train split into each tool's training layout, written to scratch directories. (exporters do not exist yet)

python -m poseinterface.export --tool dlc --split Train/<Project> --out work/dlc

3. Train and infer (each tool using its own API/CLI inside its env)

uv run --project envs/<tool> python envs/<tool>/train.py --data work/<tool> --out models/<tool>
uv run --project envs/<tool> python envs/<tool>/infer.py --model models/<tool> --videos Test/<Project>/... --out preds/<tool>

4. Convert predictions back to PoseMark with predictions_to_posemark and compare (main env)