---
language: en
license: cc-by-nc-sa-4.0
task_categories:
- image-to-image
tags:
- remote-sensing
- underwater
- image-restoration
---
# KAUSTics

Semi-synthetic paired through-water UAV restoration dataset from32 source images, with1280 calm-water reference patches and12800 distorted views. References are calm-water appearance targets, not independently verified water-free geometry or reflectance.

## Availability and files

The complete prepared release includes both `original/` (original patch bytes, nominally1000×1000) and `rgb256/` (the exact RGB256 cache used for final experiments). No patch regeneration or additional rendering was performed. One original degraded view is1344×1344 rather than1000×1000; the256 cache standardizes all images. If an initial public upload contains only256, this paragraph must be updated to reflect that partial availability.

`manifest_256.json` and `manifest_original.json` pair each reference with ten views: caustics configurations1/2 ×base, blur, color1, color2, color3. Blender supplies caustics/refraction effects; blur and colour modifications are empirical proxies for turbidity and absorption, not measured depth/scattering coefficients. No calibrated depth labels are provided. `source_assignments.csv` and the manifests define the exact partition.

| Split | UAV source IDs | Reference patches | Distorted views |
|---|---:|---:|---:|
| Train |26|1040|10400|
| Validation |3|120|1200|
| Test |3|120|1200|

All patches and variants from a source remain together. Source IDs are disjoint; geographic footprint independence has not been independently established. Do not repartition individual patches or variants randomly. Reference pixels are repeated across degraded views; count each reference once when reporting dataset area.

Load with the KAUSTics-Net code’s `Pairs` dataset. Training samples2080 base pairs per epoch, replacing the input with an augmentation variant with probability0.6. Validation/test enumerate all degraded views. Training and evaluation use256RGB images normalized to[-1,1]. Original and256 images are alternative representations of the same patches, not independent extra samples.

## License

The dataset is licensed under [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/).

## Citation

If you use this release in your research, please cite our [arXiv preprint](https://arxiv.org/abs/2611.03120).

```bibtex
@article{scilla2026kausticsnet,
  title   = {{KAUSTics-Net}: Diffusion-Based Through-Water {UAV} Restoration with a Semi-Synthetic Dataset},
  author  = {Scilla, Dario and Lopez, Omar A. and Johansen, Kasper and Nieuwenhuis, Brian O. and McCabe, Matthew F.},
  journal = {arXiv preprint arXiv:2611.03120},
  year    = {2026},
  doi     = {10.48550/arXiv.2611.03120},
  url     = {https://arxiv.org/abs/2611.03120}
}
```