Roadmap

anytrack is under active development. Versions follow Semantic Versioning; the 0.x line means the API/CLI may change between minor versions until 1.0.0.

0.2.x — classical tracking pipeline (current)

Feature-complete and validated:

• ROI detection → per-ROI background modelling → detection → Kalman-gated tracking → kinematics → Parquet/CSV.
• Fast decode-once streaming path, hardware decode, and cross-video batch concurrency with a live progress display.
• Odor-port detection (circular-edge matched filter) + fly→port distance.
• QC reports, input validation, and interactive debug GUIs.

Patch releases (0.2.1, …) carry bug fixes and refinements.

0.3.0 — pose estimation (next)

The focus of the next minor release is per-fly keypoint pose estimation via SLEAP, turning centroid tracks into body-part trajectories (head, thorax, abdomen tip, wings). The scaffolding is already in place and will be finished and integrated end-to-end:

• Label — anytrack label samples frames and seeds keypoints from the tracked ellipse; saves a resumable store, exportable to .slp.
• Train — anytrack train fits a centered-instance model (sleap-nn + torch).
• Infer — anytrack run --pose runs keypoint inference on tracked crops (pose_* settings: pose_backend, sleap_model_path, pose_device, pose_every_n, pose_batch_size, pose_conf_min, …).

The pose backend is an optional dependency (see Installation), so the classical pipeline stays dependency-light.

Beyond

1.0.0 will mark a stabilized API/CLI once pose is integrated and the interfaces have settled. Multi-fly-per-arena assignment and additional assay readouts are candidates for later 0.x releases.

Have a request? Open an issue on GitHub.