CornWeed blind reference-annotation audit: data and reproducibility code
Authors/Creators
Description
This record contains the data and code for the study "Same predictions, different references: a blind re-annotation audit of maize–weed detector evaluation" (submitted for publication).
Sixty images from the CornWeed dataset were re-annotated independently by two annotators blind to the original annotation, and disagreements among the pooled boxes were adjudicated by a third person blind to each box's source. The analysis plan, interpretation criteria and scoring parameters were frozen before any scoring.
The record includes the resulting reference annotation (L1), both annotators' independent annotations, anonymised adjudication records, the frozen rules and criteria, fixed detector predictions, split files and training configurations, and the code for candidate matching and grouping, reference construction, scoring and the blind adjudication interface. A full CPU rerun from this package reproduces all reported point estimates, intervals and classifications.
Image pixels are not redistributed; please obtain the images from the original CornWeed dataset (https://doi.org/10.5281/zenodo.7961764). Model weights, training logs and adjudication task packages are not included.
Licences: data and documentation under CC BY 4.0; project code under MIT. See README.md for details.