Synthetic CR-39 dataset
Perfect-ground-truth synthetic imagery for unit tests and a robustness benchmark — without claiming a physical detector response.
Origin & generation
Procedurally generated by alphatrack_core.synthetic: dark elliptical "track" pits on a non-uniform field, with randomized morphology, acquisition effects and artifacts.
Synthetic / real classification
100% synthetic. Stamped synthetic=true at dataset, plate, image, detection, metric and report levels; never mixed with real metrics.
Licence
CC-BY-4.0 (generator is MIT). No third-party imagery.
Plate / image counts
Deterministic per seed; the published benchmark uses a 7 / 2 / 3 plate split (train / val / test).
Transformations
Illumination gradient, vignetting, focus blur, sensor noise; artifacts: scratches, dust, hot pixels.
Splits
Plate-level (physical plate / acquisition session). Tile-level splits are refused to prevent texture leakage.
Known biases
Shapes are plausible but not a calibrated CR-39 response; distribution differs from real plates (the synthetic-to-real gap).
Quality checks & curator
Focus, saturation, illumination-uniformity and calibration checks; a locked benchmark subset is curator-controlled.
Full card in source: docs/data-cards/synthetic-cr39.md.