Trained checkpoints and per-fold predictions for "Underwater Corrosion Detection with YOLO26: Comparing Model Scaling, Ensemble Fusion, and Test-Time Augmentation"

Trained YOLO26 checkpoints and per-fold predictions of the revised study "Underwater Corrosion Detection with YOLO26: Comparing Model Scaling, Ensemble Fusion, and Test-Time Augmentation" (Ocean Engineering, manuscript OE-D-26-08406). The code is in https://doi.org/10.5281/zenodo.23337632. Files: checkpoints_main_study.zip: the 90 fold models of the grouped nested cross-validation (YOLO26 Nano, Medium and Large; direct and sequential transfer; 3 repeats x 5 outer folds). checkpoints_improvement_variants.zip: the 75 fold models of the five improvement variants (15 each). checkpoints_trashcan_and_aerial.zip: the 3 YOLO26-Medium models retrained on TrashCan 1.0 (3 seeds) and the 3 in-air corrosion pre-training models (Nano, Medium, Large) used by sequential transfer. predictions_per_fold.zip: per-fold predictions of the main study (grouped run and the ungrouped ablation, stored down to confidence 0.001), out-of-fold detections, SAM 3 outputs, and the predictions and masks of the improvement round. ext_data_derived_labels.zip: corrosion polygons of the 261 pool images and nine-class boxes and polygons of the pool and the 1,591 extra LIACI images, derived from the LIACI annotations; the images themselves must be taken from the LIACI release. results_rev2_artifacts.zip (added in version 3.0): out-of-fold predictions, training curves and logs of the 45 YOLO26-Medium models retrained with a temporal buffer of 5, 15 and 30 frame indices, and the SAM 3 masks and boxes of the second-revision re-evaluation. The buffer checkpoints are not included; the released splits and script regenerate them. SHA256SUMS.txt: SHA-256 checksums of the checkpoint archives and of results_rev2_artifacts.zip. Each checkpoint is stored with the args.yaml and results.csv of its run. Load the checkpoints with ultralytics==8.4.5. License: the models were trained on LIACI, which is available for non-commercial use only, so these files are released under CC BY-NC 4.0. They were trained with Ultralytics (AGPL-3.0).