pytc/trackingBench

WormTrack-Verify — v0.1.0-alpha.1

A public development pilot for testing anatomical verification of anonymous neuron trajectories in C. elegans. This is usable data and executable evaluation machinery. The full initial benchmark and release gates A/B/C are not complete. There is no frozen held-out test set in this release.

Motivation and scientific hypothesis

Tracking errors contaminate neural activity measurements. Eutely supports persistent neuronal populations in adult hermaphrodites; it does not imply 302 visible detections or immutable positions. The hypothesis is that an independently estimated, restricted deformation model can predict neuronal positions with uncertainty sufficiently smaller than local inter-neuron spacing to expose identity errors. Four planar eigenworm modes do not establish a four-dimensional model of 3D internal tissue motion. The benchmark must allow this hypothesis to fail.

Inventory and actual included material

Component	Observed count / status
Recording-level registry	264
Original WormID source assets	118
ASCENT human-corrected observations	105599
ASCENT anonymous tracks	96
Hosted imaging excerpts	44 volumes across four recordings
Reference point clouds	2
Independent behavior summaries	6,400 samples; one SF acquisition
Measured centerlines / eigenworm coefficients	0; unavailable
Frozen train/validation/test	Not released; development only

The five standardized tracking tables cover ASCENT, ZephIR, a NeRVE-derived recording redistributed in ZephIR, Targettrack, and a 16-frame WormID EY excerpt. Only ASCENT is currently eligible for the trusted tracking and corruption pilot. ZephIR scorer codes and Targettrack mask provenance are retained; unresolved annotations are unknown, not ground truth. The archive also contains a mouse experiment, which is excluded.

All 36 WormID tracking acquisitions and 38 excluded SF recordings are mapped at inventory level. NP/HL structural recordings are not presented as dynamic tracking. Assembly report distinguishes inventoried assets, inspected data, validated tables, and trusted correspondence labels. Source audit records blocked resources.

Tracking versus verification

tracking_gt isolates association on fixed source-supported objects and preserves the original WormID reference-frame protocol for future reproduction. anatomical_verification tests labeled identity corruptions. anatomical_consistency contains consistency diagnostics with incomplete labels. These protocols are distinct from train/validation/test partitions.

The pilot provides two untrained oracle association baselines (Euclidean nearest neighbor and Hungarian assignment), six controlled error types, clean controls, and a fixed-reference-distance consistency baseline. Its threshold is a declared 5 micrometers, chosen as a development demonstration, not calibrated to a guaranteed false-acceptance target. Reported results are descriptive development diagnostics. No full-image tracking, HOTA/IDF1, independent deformation fitting, or cross-laboratory generalization claim is made.

Data formats and coordinates

Structured data are recording-partitioned Parquet. Original annotation HDF5 files and a small original structural NWB are retained where allowed. Imaging pilots use lossless chunked HDF5 with one CZYX dataset per hosted frame; source intensities are unnormalized. The logical volume axis order is TCZYX. This HDF5 representation is not advertised as OME-Zarr/NGFF.

Point coordinates are named XYZ. Native voxel coordinates and transforms are preserved. Physical fields are null when calibration is unavailable. ASCENT spacing is (0.243, 0.243, 1.5) micrometers in XYZ, from the source record. WormID spacing comes from the inspected NWB. ZephIR normalized coordinates are converted through the source pixel-center convention; 17 out-of-bounds observations remain present with validity=false. Targettrack centroids remain in its source-processed image coordinates with an unresolved transform direction.

Ground-truth provenance

ASCENT tracks were initialized with ZephIR and manually proofread/corrected by the source authors. This is human-corrected algorithm output, not independent de novo dense annotation. Source predictions live separately under evaluation/source_predictions/ because their detections differ from oracle objects. Anonymous track IDs are sufficient; neuronal names are optional and are not used by the verifier.

The WormID EY inspected source has 962 image frames and 961 segmentation frames. Only the available observations are published. No coordinate clamping, identity interpolation, or missing annotation fabrication is performed.

Anatomy and posture

ASCENT's proofread first frame and a same-animal NeuroPAL structural cloud are included. Neither is a universal anatomical atlas. Unknown reference uncertainty is null. SF measured behavior summaries retain source timestamps and empty/unknown units. They are not centerlines, body coordinates, or a validated 3D deformation map. The client raises an explicit missing-component error for unsupported posture or imaging frames.

Splits and duplicates

All included examples are assigned to development; train/validation/test are empty and no test performance is claimed. Original WormID fold values are preserved verbatim as a reproduction reference, not interpreted as this benchmark's partitions. The original split CSV has 120 rows and 118 unique filenames; two repeated structural entries are preserved in that reference table.

Shared DANDI subject identifiers, particularly SK1 and DANDI 001623, are flagged as possible overlaps. Common animals/acquisitions and derivatives must stay together; unresolved overlap cannot be used to establish split disjointness. The NeRVE excerpt in ZephIR is not counted as an additional independent source acquisition. Test labels, when frozen in a later release, cannot calibrate corruptions or verification thresholds.

Access examples

from datasets import load_dataset
# Pin a commit SHA or the alpha tag after publication.
registry = load_dataset("pytc/trackingBench", "recordings", revision="v0.1.0-alpha.1")
tracks = load_dataset("pytc/trackingBench", "tracks", revision="v0.1.0-alpha.1", streaming=True)
corruptions = load_dataset("pytc/trackingBench", "perturbations", revision="v0.1.0-alpha.1", streaming=True)
print(next(iter(tracks["development"])))

The named configurations are recordings, detections, tracks, posture, anatomy, perturbations, and verification. The posture configuration currently contains behavior summaries, with explicit component_type, and zero measured centerlines. Loading tables does not download microscopy volumes. The recording registry includes inventory-only acquisitions; hosted_tracks and hosted_volume distinguish available components.

For the bundled selective Python client, install this repository's small code subset:

from huggingface_hub import snapshot_download
import subprocess, sys
code = snapshot_download("pytc/trackingBench", repo_type="dataset", revision="v0.1.0-alpha.1",
                         allow_patterns=["src/**", "pyproject.toml", "README.md", "requirements.lock.txt"])
subprocess.check_call([sys.executable, "-m", "pip", "install", code])
from cell_tracking.data.wormtrack_verify import BenchmarkClient
client = BenchmarkClient(revision="v0.1.0-alpha.1")
annotations = client.load_tracks("ascent-opterra")
volume = client.load_volume("ascent-opterra", frame=0)  # Small hosted excerpt only; CZYX.

Use the executable pilot to reproduce the corruption and metric demonstration. Files larger than the client's 128 MiB budget require an explicit increased budget; no function silently downloads the benchmark or a multi-GB original movie. Original assets resolve through immutable DANDI asset IDs and versioned Zenodo records in metadata/upstream_assets.json. External assets are not Hub-hosted movies.

Licensing and attribution

The original empty repository displayed MIT. That does not license third-party data. Included DANDI and Zenodo data are CC BY 4.0 per their release metadata. Benchmark helper code is MIT; source data retain their own licenses and attribution. Resources with unresolved redistribution rights remain references only. See per-source licensing, catalog, and citations.

Limitations and biases

This pilot has only one trusted proofread tracking specimen, no measured centerlines, no independently calibrated deformation uncertainty, and no held-out evaluation. Unknown scorer codes are not promoted to manual ground truth. Fixed-reference residuals include legitimate tissue motion and are permutation-sensitive only because reference correspondences remain fixed. A static unordered point cloud cannot detect identity swaps. Acceptance by a geometric heuristic is not proof of biological identity.

Coverage and selective risk use the full declared observation population; rejecting everything yields zero coverage. Sources differ in optics, visibility, annotation construction, and acquisition restrictions. The 302-neuron assumption is not universal across sex/stage and is never used to force observed counts.

Release history

• v0.1.0-alpha.1: development pilot; source inventory, five standardized tracking tables, imaging excerpts, two reference clouds, measured behavior summaries, controlled corruptions, and reference baseline outputs. Full initial release gates remain open.

Each release has a SHA256 assembly manifest, an upstream asset manifest, dependency pins, and machine-readable validation results. No model weights or externally pretrained models are used by these pilot baselines.