dar96io/Sea-Undistort-evaluation

Sea-Undistort: KAUSTics-Net experiment partition

This release reorganizes the original Sea-Undistort data into the fixed training, validation, and test splits used in our KAUSTics-Net preprint. Original creators: Maximilian Kromer, Panagiotis Agrafiotis and Begüm Demir (2025). Original dataset: https://huggingface.co/datasets/maxkromer/Sea-Undistort ; DOI https://doi.org/10.5281/zenodo.15639838 ; project https://www.magicbathy.eu/Sea-Undistort.html .

Retain the upstream CC BY-NC-SA4.0 terms and attribution: https://creativecommons.org/licenses/by-nc-sa/4.0/ . Preparation reorganizes original image bytes into fixed scene partitions and provides a pairing manifest. No pixels were modified during this release preparation. The code converts toRGB and bicubically resizes to256 on loading.

manifest.json contains840 train,240 validation and120 test scene IDs, without scene overlap. Train/validation inputs are all-effects. Test includes three conditioning variants of the same120 scenes: all-effects, no-waves and no-sunglint. All reference targets are original GT files; no-waves/no-sunglint are inputs and must not be substituted for GT. Files are under images/<split>/<gt|all_effects|no_waves|no_sunglint>/.

The payload contains2640 image files:1200 GTs,1200 all-effects inputs,120 no-waves inputs and120 no-sunglint inputs. Other original dataset metadata and unused input variants for train/validation remain available from the original release. Checksums are included in the pairing manifest.

Please cite Kromer, Agrafiotis and Demir, Sea-Undistort: A Dataset for Through-Water Image Restoration in High Resolution Airborne Bathymetric Mapping, IEEE Geoscience and Remote Sensing Letters22 (2025), DOI10.1109/LGRS.2025.3601239 when using Sea-Undistort.