RealRain-1k: Real-World Single-Image Deraining Dataset (Unofficial Mirror)

Unofficial redistribution of RealRain-1k, the real-world rain/rain-free image dataset from Li et al. (arXiv 2022), packaged for direct use with ClearView's dataset pipeline.

Disclaimer

This repository is not an official release of RealRain-1k.

RealRain-1k was created by Wei Li, Qiming Zhang, Jing Zhang, Zhen Huang, Xinmei Tian, and Dacheng Tao. This repository does not claim ownership of any images or metadata, and attributes the dataset to its original creators. (We describe them as "creators" rather than "copyright holders" deliberately — copyright in academic datasets can rest with an author's university, a funding body, or otherwise, under arrangements we have no way to verify from the outside. The official repository's MIT license, discussed under License below, is what actually governs redistribution here.)

This repository exists to provide a directly downloadable mirror — the official dataset is distributed via OneDrive/Google Drive folder links rather than a versioned, scriptable source, which is inconvenient for reproducible pipelines and CI.

Single-hop provenance. Unlike this collection's SPA-Data mirror, this one does not appear to involve a subsampled or third-party intermediate copy: every split's pair count (784/112/224 per density track) matches the official paper's reported totals exactly, with no evidence of subsampling. We're not able to cryptographically verify it is byte-identical to the official OneDrive/Google Drive release, but the counts line up exactly with what's published.

---

Dataset Description

RealRain-1k is a real-world (not synthetically rendered) single-image deraining benchmark. Rain/clean pairs were "automatically generated from a large number of real-world rainy video clips through a simple yet effective rain density-controllable filtering method" (per the paper), yielding two parallel tracks at different rain densities:

• RealRain-1k-H (Heavy): denser, more visually severe rain streaks.
• RealRain-1k-L (Light): the same underlying scene pool, filtered at a lower rain-density threshold.

Both tracks share the same per-scene numeric IDs, high resolution (variable — not fixed-size patches; images range from roughly 550×550 up to 1920×1080 in this copy), background diversity, and strict spatial alignment between rain and clean frames. The authors also released a companion synthetic dataset, SynRain-13k, generated by extracting the real rain-streak layers and compositing them onto natural images — not included in this mirror, which covers RealRain-1k only.

---

Changes from the Official Release

• Removed 9 .DS_Store files (macOS Finder metadata, present in every split of the source copy — not part of the dataset).
• Removed 1 duplicate file: RealRain-1k-L/test/target/1101 [conflicted].png, a byte-identical duplicate of 1101.png left behind by a cloud-sync conflict (Dropbox/OneDrive-style naming) in the source copy. 1101.png itself is untouched and present.
• No images added, modified, or otherwise removed. No relabeling. No further subsampling.
• Directory layout preserved exactly as received ({H,L}/{train,validation,test}/{input,target}).

---

Dataset Structure

realrain-1k/
├── README.md
├── banner.jpg
├── RealRain-1k-H/
│   ├── train/{input,target}/          # 784 pairs
│   ├── validation/{input,target}/     # 112 pairs
│   └── test/{input,target}/           # 224 pairs
└── RealRain-1k-L/
    ├── train/{input,target}/          # 784 pairs
    ├── validation/{input,target}/     # 112 pairs
    └── test/{input,target}/           # 224 pairs

input/target filenames share identical stems (e.g. input/104.png ↔ target/104.png), so this mirror works directly with a generic paired-image loader — no custom filename-matching logic required (unlike SPA-Data's rain-{id}/norain-{id} prefix mismatch).

Track	Split	Pairs	Size on disk
H (Heavy)	train / validation / test	784 / 112 / 224	~2.4 GB
L (Light)	train / validation / test	784 / 112 / 224	~2.3 GB
Total	—	2,240	~4.6 GB

---

Usage with ClearView

ClearView doesn't need a dedicated parser class for this dataset — since input/target filenames match exactly, the generic ImagePairDataset handles it directly:

from huggingface_hub import snapshot_download
from clearview.data import ImagePairDataset, get_train_transforms, get_val_transforms

data_dir = snapshot_download(repo_id="dronefreak/RealRain-1k", repo_type="dataset")

# Pick a track: RealRain-1k-H (heavy) or RealRain-1k-L (light)
track = f"{data_dir}/RealRain-1k-H"

train_ds = ImagePairDataset(
    rainy_dir=f"{track}/train/input",
    clean_dir=f"{track}/train/target",
    transform=get_train_transforms(crop_size=(256, 256)),
)
val_ds = ImagePairDataset(
    rainy_dir=f"{track}/validation/input",
    clean_dir=f"{track}/validation/target",
    transform=get_val_transforms(),
)

rainy, clean = train_ds[0]

Or directly via the training CLI (--dataset-type pair with explicit path overrides):

clearview-train \
  --data-dir <path-to-downloaded-snapshot>/RealRain-1k-H \
  --dataset-type pair \
  --train-rainy train/input --train-clean train/target \
  --val-rainy validation/input --val-clean validation/target \
  --model restormer --batch-size 8 --crop-size 256 --epochs 100 \
  --output-dir ./runs/realrain1k_h_restormer

Swap RealRain-1k-H → RealRain-1k-L for the light-rain track. Evaluate against the held-out test split the same way via clearview-evaluate --dataset-type pair --rainy-dir test/input --clean-dir test/target.

---

Dataset Sources

Original Paper

Toward Real-world Single Image Deraining: A New Benchmark and Beyond

Wei Li, Qiming Zhang, Jing Zhang, Zhen Huang, Xinmei Tian, Dacheng Tao

arXiv preprint, 2022.

• arXiv: https://arxiv.org/abs/2206.05514

Official Resources

• Code + Dataset: https://github.com/hiker-lw/RealRain-1k
• OneDrive: https://1drv.ms/u/s!AimBgYV7JjTlgg1MmR2tfBPW1Egh?e=rUNw3m
• Google Drive: https://drive.google.com/drive/folders/1rk7jdBZifNe_OKJ6j-0Ne8gjYYIg6Qc5

---

Attribution

All credit for the dataset belongs entirely to the original authors: Wei Li, Qiming Zhang, Jing Zhang, Zhen Huang, Xinmei Tian, and Dacheng Tao.

If you use this dataset in your research, please cite the original publication below.

---

License

The official RealRain-1k repository ships an explicit MIT License, the most permissive of the licenses found across the datasets in this mirror collection — it allows unrestricted use, modification, and redistribution (including commercial), provided the copyright notice is retained.

This repository is distributed under the same terms: MIT.

---

Citation

If you use this dataset, please cite:

@article{li2022toward,
  title={Toward Real-world Single Image Deraining: A New Benchmark and Beyond},
  author={Li, Wei and Zhang, Qiming and Zhang, Jing and Huang, Zhen and Tian, Xinmei and Tao, Dacheng},
  journal={arXiv preprint arXiv:2206.05514},
  year={2022}
}

---

Acknowledgements

We sincerely thank Wei Li, Qiming Zhang, Jing Zhang, Zhen Huang, Xinmei Tian, and Dacheng Tao for creating and publicly releasing this valuable real-world deraining benchmark.