SPA-Data: Real-World Single-Image Deraining Dataset (Unofficial, Subsampled Mirror)

Unofficial, subsampled redistribution of SPA-Data, the real-world rain/rain-free image dataset from SPANet (Wang et al., CVPR 2019), packaged in a directory layout directly consumable by ClearView's SPADataDataset parser.

Disclaimer

This repository is not an official release of SPA-Data.

SPA-Data was created by Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson W.H. Lau. 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 BSD-3-Clause license, discussed under License below, is what actually governs redistribution here.)

This repository exists for two purposes:

1. To provide a directly downloadable mirror. The official training set is distributed as a 12-part split archive (real_world.zip.00–.11, ~44GB) that must be manually concatenated and unzipped — inconvenient for quick experimentation.
2. To provide a subsampled training set (6,385 of the official ~29,500 pairs, ~22%) sized for fast iteration on a single workstation GPU, while leaving the full official validation/test split (1,000 pairs) untouched, so evaluation numbers on this mirror's val split remain comparable to published results.

Two-hop provenance. This redistribution is not sourced directly from the official 44GB multi-part archive. It is sourced from a pre-subsampled copy whose directory layout and exact training-set size (6,385 pairs) match a Kaggle re-upload also called "SPA-Data" (leftthomas/spadata), which is referenced by name in some downstream repositories in the deraining literature. We were not able to independently verify that Kaggle listing's own stated license via automated retrieval while preparing this card — see License below for what this mirror's terms are actually based on.

---

Dataset Description

SPA-Data is a real-world (not synthetically rendered) single-image deraining benchmark, introduced alongside the SPANet architecture. The authors collected 170 real rain videos — 84 scenes captured on iPhones plus 86 scenes sourced from stock-footage libraries (StoryBlocks) and YouTube — and used a semi-automatic method combining temporal priors with human supervision to reconstruct a clean, rain-free reference frame for each rainy sequence, yielding strictly spatially-aligned rain/clean pairs. This is the standard reason SPA-Data is used as a real-world generalization check for models trained on synthetic rain (Rain13K, DDN-Data, etc.) — synthetic rain streaks don't fully capture the shape, direction, and intensity variation of real rain.

Officially: ~29.5K rain/rain-free training pairs, plus a held-out 1,000-pair test split.

This mirror: 6,385 training pairs (subsampled) + the full 1,000-pair official validation/test split.

---

Changes from the Official Release

The transformation chain has two hops, each scoped narrowly:

1. Official 44GB archive → subsampled Kaggle-style copy (not performed by this repository)

An intermediate source (see Disclaimer) subsampled the official ~29,500-pair training set down to 6,385 pairs. We do not know the exact selection method used (random subsample vs. scene-stratified — not documented anywhere we could find). If your use case is sensitive to sampling bias in the training set, be aware of this. The validation/test split was not subsampled — it retains all 1,000 official pairs.

2. Subsampled copy → this repository

• No images added, removed, or modified.
• No labels/pairings changed.
• No further subsampling — training set is redistributed exactly as received (6,385 pairs).
• Directory layout preserved as-is (rgb_reconstruction/{rain,norain} under train//val/), since it already matches the naming convention documented in the official SPANet repo.

Note on resolution: train/ images are 256×256 patches; val/ images are 512×512 patches. This matches the crop convention used in the official release and is preserved unchanged.

---

Dataset Structure

spa-data/
├── README.md
├── banner.jpg
├── train/
│   └── rgb_reconstruction/
│       ├── rain/         # rain-{id}.png   (6,385 images, 256x256)
│       └── norain/       # norain-{id}.png (6,385 images, 256x256)
└── val/
    └── rgb_reconstruction/
        ├── rain/         # rain-{id}.png   (1,000 images, 512x512)
        └── norain/       # norain-{id}.png (1,000 images, 512x512)

Rain and clean images share a common numeric ID but different filename prefixes (rain-{id}.png / norain-{id}.png), not identical stems — this is why ClearView ships a dedicated SPADataDataset parser rather than relying on the generic ImagePairDataset (which expects exact-stem matches).

Split	Pairs	Resolution	Size on disk
train	6,385	256×256	~806 MB
val	1,000	512×512	~451 MB
Total	7,385	—	~1.3 GB

---

Usage with ClearView

This dataset is designed to be used directly with ClearView, an open-source PyTorch framework for image deraining.

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

data_dir = snapshot_download(repo_id="dronefreak/SPA-Data", repo_type="dataset")

train_ds = SPADataDataset(root_dir=data_dir, split="train", transform=get_train_transforms(crop_size=(256, 256)))
val_ds = SPADataDataset(root_dir=data_dir, split="val", transform=get_val_transforms())

rainy, clean = train_ds[0]

Or directly via the training CLI:

clearview-train \
  --data-dir <path-to-downloaded-snapshot> \
  --dataset-type spa-data --train-split train --val-split val \
  --model restormer --batch-size 8 --crop-size 256 --epochs 100 \
  --output-dir ./runs/spa_data_restormer

SPADataDataset auto-detects the rgb_reconstruction/{rain,norain} layout and matches pairs by numeric ID after stripping the rain-/norain- prefixes — no path overrides needed beyond --train-split/--val-split.

---

Dataset Sources

Original Paper

Spatial Attentive Single-Image Deraining with a High Quality Real Rain Dataset

Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, Rynson W.H. Lau

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019.

• arXiv: https://arxiv.org/abs/1904.01538
• CVPR Open Access (PDF): https://openaccess.thecvf.com/content_CVPR_2019/papers/Wang_Spatial_Attentive_Single-Image_Deraining_With_a_High_Quality_Real_Rain_CVPR_2019_paper.pdf

Official Resources

• Code + Dataset: https://github.com/stevewongv/SPANet
• Official train/test dataset READMEs: Train_Dataset_README.md, Test_Datset_README.md

Possible Intermediate Source (unverified)

• Kaggle: https://www.kaggle.com/datasets/leftthomas/spadata — directory layout and 6,385-pair training-set size match this mirror; license field could not be independently confirmed via automated retrieval while preparing this card.

---

Attribution

All credit for the dataset belongs entirely to the original authors: Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson W.H. Lau.

If an intermediate Kaggle re-upload was in fact the direct source of this subsampled copy, credit for the subsampling also belongs to that contributor (see above) — we could not fully confirm this chain, and note it here in the interest of transparency rather than certainty.

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

---

License

The official SPANet repository ships an explicit BSD 3-Clause License (License.txt), and its README states: "If you use this code or our dataset (including test set), please cite" the paper below — i.e. the dataset is explicitly covered by, and offered for use under, the same repository license as the code, with no non-commercial restriction and no access-request gate.

Key BSD-3-Clause conditions: redistributions must retain the copyright notice, the license conditions, and the disclaimer; the authors' names may not be used to endorse derived products without prior written permission; the software/data is provided with no warranty.

A note on verification. This determination is based on the official SPANet GitHub repository's license file, which we did retrieve and confirm directly. The possible intermediate Kaggle source's own stated license (if different) could not be independently confirmed — see Dataset Sources. If you require certainty for commercial use, we recommend cross-checking the official repository directly.

This repository is distributed under the same terms as the official source: BSD-3-Clause.

---

Citation

If you use this dataset, please cite:

@InProceedings{Wang_2019_CVPR,
  author = {Wang, Tianyu and Yang, Xin and Xu, Ke and Chen, Shaozhe and Zhang, Qiang and Lau, Rynson W.H.},
  title = {Spatial Attentive Single-Image Deraining with a High Quality Real Rain Dataset},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2019}
}

---

Acknowledgements

We sincerely thank Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson W.H. Lau for creating and publicly releasing this valuable real-world deraining benchmark, and the community contributor(s) behind the intermediate subsampled copy this repository redistributes.