CSD: Comprehensive Snow Dataset (Unofficial Mirror)

Unofficial redistribution of CSD (Comprehensive Snow Dataset), the synthetic single-image desnowing dataset from Chen et al. (ICCV 2021), packaged for direct use with ClearView's dataset pipeline. This is ClearView's first desnowing dataset, distinct from its existing deraining-focused mirrors.

Disclaimer

This repository is not an official release of CSD.

CSD was created by Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I-Hsiang Chen, Jian-Jiun Ding, and Sy-Yen Kuo. 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.)

---

Dataset Overview

CSD is a synthetic single-image desnowing benchmark, built by compositing rendered snow particles (varying transparency, size, and position, with Gaussian blur applied for realistic focus falloff) onto clean photographs, alongside a mild haze layer to better approximate the look of a real snowfall scene rather than isolated falling-snow streaks. Each pair also ships a binary snow mask locating exactly where the synthetic snow was composited, useful for auxiliary supervision (e.g. a mask-prediction loss) even though ClearView's own training pipeline doesn't currently consume it.

• Training: 8,000 pairs.
• Testing: 2,000 pairs.
• Resolution: 640×480, .tif (LZW-compressed, lossless).

---

Changes from the Official Release

None beyond repackaging. No images added, removed, or modified. No relabeling. The Train/{Gt,Mask,Snow} and Test/{Gt,Mask,Snow} directory layout matches the official archive's structure exactly, this is a full mirror, not a subsample.

---

Dataset Structure

csd/
├── README.md
├── banner.jpg
├── Train/
│   ├── Gt/      # {id}.tif — 8,000 clean images
│   ├── Snow/    # {id}.tif — 8,000 synthetic snow images
│   └── Mask/    # {id}.tif — 8,000 binary snow masks
└── Test/
    ├── Gt/      # {id}.tif — 2,000 clean images
    ├── Snow/    # {id}.tif — 2,000 synthetic snow images
    └── Mask/    # {id}.tif — 2,000 binary snow masks

Gt/Snow/Mask share identical filenames within each split (e.g. 1000.tif appears in all three), exact-stem matching, no custom parser needed.

Split	Pairs	Resolution	Format
Train	8,000	640×480	.tif
Test	2,000	640×480	.tif
Total	10,000

---

Usage with ClearView

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

CSD's Gt/Snow folders share identical filenames, so ClearView's generic ImagePairDataset works directly, no dedicated parser needed, unlike SPA-Data or DDN-Data. The one thing to set explicitly is the file extension, since CSD ships .tif rather than ImagePairDataset's .png/.jpg/.jpeg default:

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

data_dir = snapshot_download(repo_id="dronefreak/CSD", repo_type="dataset")

train_ds = ImagePairDataset(
    rainy_dir=f"{data_dir}/Train/Snow",   # "rainy_dir" is just the degraded-image arg name; snow works the same way
    clean_dir=f"{data_dir}/Train/Gt",
    transform=get_train_transforms(crop_size=(256, 256)),
    extensions=(".tif",),
)
test_ds = ImagePairDataset(
    rainy_dir=f"{data_dir}/Test/Snow",
    clean_dir=f"{data_dir}/Test/Gt",
    transform=get_val_transforms(),
    extensions=(".tif",),
)

snowy, clean = train_ds[0]

A caveat on the clearview-train CLI specifically: --dataset-type pair instantiates ImagePairDataset without passing through an extensions argument (see clearview/scripts/train.py's _build_dataset helper), so it's hardcoded to the .png/.jpg/.jpeg default and won't find this dataset's .tif files as-is. Until a --extensions flag is wired through the CLI, use the Python API above directly (which does accept extensions=(".tif",)), or convert the .tif files to .png first.

---

Dataset Sources

Original Paper

ALL Snow Removed: Single Image Desnowing Algorithm Using Hierarchical Dual-Tree Complex Wavelet Representation and Contradict Channel Loss

Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I-Hsiang Chen, Jian-Jiun Ding, Sy-Yen Kuo

IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 4196–4205.

• CVF Open Access (PDF): https://openaccess.thecvf.com/content/ICCV2021/papers/Chen_ALL_Snow_Removed_Single_Image_Desnowing_Algorithm_Using_Hierarchical_Dual-Tree_ICCV_2021_paper.pdf

Official Resources

• Code + Dataset: https://github.com/weitingchen83/ICCV2021-Single-Image-Desnowing-HDCWNet
• Dataset download (Google Drive): linked from the official repository's README

---

Attribution

All credit for the dataset belongs entirely to the original authors: Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I-Hsiang Chen, Jian-Jiun Ding, and Sy-Yen Kuo.

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

---

License

The official repository ships an explicit MIT License. This repository is distributed under the same terms.

MIT is a permissive license: you're free to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies, provided the original copyright notice and permission notice are retained. No non-commercial restriction, no access-request gate, the cleanest licensing situation of any dataset mirrored into this collection so far.

---

Citation

If you use this dataset, please cite:

@InProceedings{Chen_2021_ICCV,
  author = {Chen, Wei-Ting and Fang, Hao-Yu and Hsieh, Cheng-Lin and Tsai, Cheng-Che and Chen, I-Hsiang and Ding, Jian-Jiun and Kuo, Sy-Yen},
  title = {ALL Snow Removed: Single Image Desnowing Algorithm Using Hierarchical Dual-Tree Complex Wavelet Representation and Contradict Channel Loss},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month = {October},
  year = {2021},
  pages = {4196-4205}
}

---

Acknowledgements

We sincerely thank Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I-Hsiang Chen, Jian-Jiun Ding, and Sy-Yen Kuo for creating and publicly releasing this dataset, and for licensing it permissively.