Reproducibility

1. Download the annotated cucumber dataset from the associated Zenodo repository (https://doi.org/10.5281/zenodo.16816441).
2. Execute the dataset preparation and split-generation sections to reproduce the fixed 15% holdout test set and the three training-validation folds.
3. Use the frozen final recipe and the random seeds 42, 123, 2024, 3407, and 777 to train the six detector configurations.
4. Evaluate each selected best.pt checkpoint on the fixed holdout test set.
5. Execute the statistical analysis section to reproduce the aggregated metrics, bootstrap intervals, paired permutation and Wilcoxon tests, Holm-adjusted p-values, class-wise results, and accuracy–efficiency analysis.