PixelShift200

Full-color-sampled, demosaicing-artifact-free 4K images for training demosaicing, denoising and super-resolution models.

_{Eight of the 109 captured scenes, rendered from the RAW previews.}

PixelShift200 is the dataset introduced in Rethinking Learning-based Demosaicing, Denoising, and Super-Resolution Pipeline (ICCP 2022). It was captured with a SONY α7R III using pixel shift technology: the camera takes four samples of the same scene, physically moving the sensor by one pixel horizontally or vertically between samples, so that R, Gr, Gb and B are all measured at every pixel location.

Because no channel is ever interpolated, the resulting images are free of demosaicing artifacts while still following the distribution of natural images as sampled by a real camera sensor. Models trained on PixelShift200 do not learn to reproduce the demosaicing artifacts that are baked into conventional RGB training sets.

• 📄 Paper: ICCP 2022 · arXiv:1905.02538
• 🌐 Project page: https://guochengqian.github.io/project/pixelshift200/
• 💻 Code: https://github.com/guochengqian/TENet

Dataset summary

Training images	200, 4K resolution
Validation images	10, 1K resolution
Cropped training patches	9,444 at 512×512
Source RAW captures	109 scenes × 4 pixel-shift frames = 436 .ARW files
Camera	SONY α7R III
Format	.mat, 4 channels — R, Gr, Gb, B
Total size	~113 GB

Repository layout

readme.txt                                    original release notes
PixelShift200/
  PixelShift200.zip                 14.7 GB   the dataset: 200 train (4K) + 10 val (1K), .mat
  train_rggb_512.zip                13.2 GB   9,444 training crops at 512×512, .mat
  deprecated_version_with_color/              superseded earlier release, kept for reproducibility
    PixelShift200_train.zip         10.2 GB
    PixelShift200_test.zip           0.5 GB
PixelShift200_RawFiles/
  ARW/                              37.4 GB   436 camera RAW files (109 scenes × 4 shifted frames)
  raw_rggb/
    train_raw_rggb/                 35.4 GB   104 full-resolution merged RGGB scenes, .mat
    test_raw_rggb/                   1.7 GB   5 held-out scenes, .mat
    preview/                         0.2 GB   109 sRGB .jpg previews of the RAW captures

Most users only need PixelShift200/PixelShift200.zip (or train_rggb_512.zip for pre-cropped patches). The PixelShift200_RawFiles/ tree is for anyone who wants to redo the RAW processing from the original camera files.

Usage

Download only what you need:

from huggingface_hub import hf_hub_download

# the main dataset (14.7 GB)
path = hf_hub_download(
    repo_id="guochengqian/PixelShift200",
    filename="PixelShift200/PixelShift200.zip",
    repo_type="dataset",
)

# or the pre-cropped 512x512 training patches (13.2 GB)
path = hf_hub_download(
    repo_id="guochengqian/PixelShift200",
    filename="PixelShift200/train_rggb_512.zip",
    repo_type="dataset",
)

Or grab a subtree with the CLI:

hf download guochengqian/PixelShift200 --repo-type dataset \
    --include "PixelShift200/*" --local-dir ./pixelshift200

Each .mat file is MATLAB v7 format, so scipy.io.loadmat reads it directly — no h5py needed:

import scipy.io as sio

d = sio.loadmat("pixelshift_10_rggb.mat")

rggb = d["rggb"]            # (H, W, 4) int16, channels R, Gr, Gb, B
print(rggb.shape, rggb.dtype, rggb.min(), rggb.max())
# (5316, 7996, 4) int16 0 16383     <- 14-bit sensor values

meta = d["metadata"][0, 0]
meta["colormatrix"]         # (3, 3) float32, camera RGB -> sRGB
meta["rgb_gain"]            # global gain
meta["red_gain"], meta["blue_gain"]   # white-balance gains

Values are unnormalised 14-bit sensor readings (0–16383). The metadata struct carries everything needed to render a scene to sRGB: apply the white-balance gains, then the color matrix, then a gamma curve.

To reproduce the 512×512 crops yourself, or to train TENet on this data, follow the data-preparation steps in the TENet repo (datasets/crop_pixelshift200.py).

How it was captured

Pixel shift takes four exposures of a static scene, translating the sensor by exactly one pixel between exposures. Combining the four frames yields a measurement of every color channel at every pixel — no Bayer interpolation anywhere in the pipeline. The trade-off is that the scenes must be static, which is why the set is 210 carefully chosen images rather than a large in-the-wild collection.

The .ARW files in PixelShift200_RawFiles/ARW/ are the untouched four-frame bursts; raw_rggb/ holds the merged four-channel results before cropping.

Citation

@inproceedings{qian2022rethinking,
  title={Rethinking Learning-based Demosaicing, Denoising, and Super-Resolution Pipeline},
  author={Qian, Guocheng and Wang, Yuanhao and Gu, Jinjin and Dong, Chao and Heidrich, Wolfgang and Ghanem, Bernard and Ren, Jimmy S},
  booktitle={2022 IEEE International Conference on Computational Photography (ICCP)},
  pages={1--12},
  year={2022},
  organization={IEEE}
}

License

Released under CC BY 4.0. Please cite the paper above if you use PixelShift200.