v0.2.1
First 0.2.x release. Highlights since v0.1.7:
Packaging & deployment
- Headless-first install: default
pip install openptv2is now the slim batch runtime (no Qt/GUI); desktop GUI viaopenptv2[gui]. - Cloud batch:
Dockerfile.cloud(free-threaded 3.14t, no GUI) + docs/cloud-batch.md with one-command install and measured timings. - Python 3.11–3.14 supported; cibuildwheel wheels for cp311–cp313 (Linux/Windows/macOS).
Tracking
- Cold-start recovery + forward/backward reciprocity post-passes.
- 3D segment tracking fixes (
full_forward_3doff-by-one, finalize). - Fix: forward 2D tracking ran one step past the sequence end.
Performance
- Correspondences ~157× faster (flat numpy arrays), target recognition ~63× (
targ_rec_fast), GIL-free BFS + parallel 4-camera detection ~2.3×.
Parameters
- YAML-first experiment discovery with explicit active-run selection.
- Pydantic v2 parameter validation.
GUI
- Fix:
openptv2-guiwith no args now prints guidance instead of crashing. - Fix: calibration residual lines no longer drawn from image origin.
Wheels and sdist are attached automatically by CI.