RESIDE-ITS: Indoor Training Set for Single Image Dehazing (Unofficial Mirror)

Unofficial redistribution of RESIDE's ITS (Indoor Training Set) subset, from Li et al.'s RESIDE benchmark (IEEE TIP 2019), packaged for direct use with ClearView's dataset pipeline. Synthetic indoor haze pairs: one clean photograph, ten hazed variants rendered against it at different scattering coefficients.

Disclaimer

This repository is not an official release of RESIDE.

RESIDE was created by Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. 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.

On licensing, please read before relying on this mirror. We looked for a formal license covering this dataset (project homepage, the official link-only GitHub repo, the paper itself) and found none for the image data. The GitHub repo carries an MIT badge, but that covers the repo's own contents, a text file of Dropbox/Baidu download links, not the images themselves. The paper's only rights language is standard IEEE copyright boilerplate covering the PDF text. This is the same situation as this collection's Rain13K, DDN-Data, RainDrop, NTIRE-Haze, and RainDS mirrors, not a stated restriction like a dataset that explicitly says "research purposes only" (which we treat differently and do not mirror, see this collection's decision on GT-RAIN). In the absence of an explicit license, we have chosen to preserve this dataset for research reproducibility while providing full attribution and making the mirror removable on request, this is a decision we're making under that uncertainty, not a claim that we've established a legal right to redistribute it. Concretely:

• Full attribution and citation to the original creators (below).
• No claim of any rights beyond hosting a copy.
• If the copyright holder, or an authorized representative, requests removal, we will comply promptly.

A note on image content. The underlying clean indoor photographs originate from the NYU Depth Dataset V2 (Silberman et al.) and the Middlebury stereo dataset, both established academic depth-estimation benchmarks. We screened a spread of samples across the full ID range (train and val) before mirroring and found empty interior scenes (basements, bathrooms, offices, bedrooms, a library), no people, faces, license plates, or embedded watermarks. This is a materially different provenance from a web-scraped photo compilation, see Snow100K's evaluation notes for the kind of risk this check is meant to catch.

If you need certainty about redistribution rights (e.g. for commercial use), contact the dataset's creators or their institution directly rather than relying on this mirror.

---

Dataset Overview

RESIDE is a large synthetic single-image dehazing benchmark built by rendering haze onto real, haze-free photographs using the atmospheric scattering model, rather than photographing actual haze. This mirror covers ITS (Indoor Training Set) only, the indoor subset used for training; RESIDE's outdoor subsets (OTS, RTTS) are not included here, see Changes from the Official Release.

Each clean image has exactly ten hazy variants, rendered at different scattering coefficients (β) and atmospheric light values (A):

Split	Clear images	Hazy images (10× clear)	Transmission maps
train	10,000	100,000	100,000
val	1,000	10,000	10,000
Total	11,000	110,000	110,000

A note on counts. The original paper states ITS as 13,990 hazy images generated from 1,399 clear images. The archive actually in wide distribution today (verified against every file in this mirror, not taken from a citation) is substantially larger, 110,000 hazy images from 11,000 clear images, as shown above. We're citing the verified file counts, not the paper's original figure, and flagging the discrepancy rather than silently picking one number: the publicly distributed ITS has apparently been expanded well beyond what the paper itself documents.

Transmission maps (trans/) are included for parity with the official release, they encode the per-pixel scene-depth term used to synthesize the haze, but are not loaded by ClearView's image-to-image training pipeline, which only needs the clean/hazy pairs.

---

Changes from the Official Release

No images added, removed, or modified, and no files renamed. This mirror covers the ITS subset only: RESIDE's much larger outdoor synthetic subset (OTS, ~313K images) and its real-world unpaired outdoor test set (RTTS) were evaluated separately and not mirrored, their images are collected from the open web rather than depth-estimation benchmarks like ITS's NYU2/Middlebury sources, the same provenance risk category that led us to skip mirroring Snow100K (see that dataset's notes in this collection).

---

Dataset Structure

reside-its/
├── README.md
├── banner.jpg
├── ITS_clear/                  # {id}.png — 10,000 clean indoor images (train)
├── ITS_haze/
│   ├── shard_000/               # {id}_01_{atm_light}.png — 10,000 hazy images (variant 1)
│   ├── shard_001/               # {id}_02_{atm_light}.png — 10,000 hazy images (variant 2)
│   ├── ...
│   └── shard_009/               # {id}_10_{atm_light}.png — 10,000 hazy images (variant 10)
├── ITS_trans/                  # same shard_000..shard_009 layout, transmission maps, unused by ClearView
└── val/
    ├── clear/                   # {id}.png — 1,000 clean indoor images
    ├── haze/                    # {id}_{variant:02d}_{atm_light}.png — 10,000 hazy images (flat)
    └── trans/                   # {id}_{variant:02d}.png — 10,000 transmission maps (flat, unused by ClearView)

The train split ships ITS_clear/ITS_haze/ITS_trans directly at the repo root (matching the official archive's own top-level folder names), while the val split nests the same three categories one level deeper under val/. Every hazy or transmission filename carries its source clear image's ID as a prefix before the first underscore (e.g. 0001_07_0.7714.png was rendered from ITS_clear/0001.png), a one-clean-to-ten-hazy broadcast rather than a 1:1 stem match, which is why ClearView ships a dedicated RESIDEITSDataset parser rather than the generic ImagePairDataset.

On the shard_NNN subdirectories: the Hugging Face Hub enforces a 10,000-files-per-directory limit, and ITS_haze/ITS_trans each have 100,000 files, ten times over. We sharded each into ten subdirectories, one per haze variant (shard_000 = variant 1, ..., shard_009 = variant 10), rather than splitting some other way, so each shard stays a clean, self-contained slice. val/haze and val/trans have exactly 10,000 files each, at the limit but not over it, so they're left flat. RESIDEITSDataset handles both layouts transparently, callers don't need to know which one a given directory uses.

---

Usage with ClearView

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

from huggingface_hub import snapshot_download
from clearview.data import RESIDEITSDataset, get_train_transforms

data_dir = snapshot_download(repo_id="dronefreak/RESIDE-ITS", repo_type="dataset")

train_ds = RESIDEITSDataset(
    data_dir,
    split="train",
    transform=get_train_transforms(crop_size=(256, 256)),
)
val_ds = RESIDEITSDataset(data_dir, split="val")

hazy, clean = train_ds[0]

Note: RESIDEITSDataset is not yet wired into the clearview-train CLI's --dataset-type choices (Python API only for now).

---

Dataset Sources

Original Paper

Benchmarking Single-Image Dehazing and Beyond

Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, Zhangyang Wang

IEEE Transactions on Image Processing (TIP), Vol. 28, No. 1, pp. 492-505, 2019.

• arXiv: https://arxiv.org/abs/1712.04143
• DOI: https://doi.org/10.1109/TIP.2018.2867951

Official Resources

• Dataset Homepage: https://sites.google.com/view/reside-dehaze-datasets/
• Download Links: https://github.com/Boyiliee/RESIDE-dataset-link

---

Attribution

All credit for the dataset belongs entirely to the original authors: Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang.

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

---

License

No formal license was located for this dataset, see the Disclaimer above for our reasoning for mirroring it anyway under that uncertainty (full attribution, reproducibility purpose, removable on request). This repository makes no claim to any rights over the dataset content beyond hosting a copy.

---

Citation

If you use this dataset, please cite:

@article{li2019benchmarking,
  author  = {Li, Boyi and Ren, Wenqi and Fu, Dengpan and Tao, Dacheng and Feng, Dan and Zeng, Wenjun and Wang, Zhangyang},
  title   = {Benchmarking Single-Image Dehazing and Beyond},
  journal = {IEEE Transactions on Image Processing},
  volume  = {28},
  number  = {1},
  pages   = {492--505},
  year    = {2019},
  doi     = {10.1109/TIP.2018.2867951}
}

---

Acknowledgements

We sincerely thank Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang for creating and publicly releasing this dataset.