SIDD-CC

SIDD-CC ("Color-Corrected") is a re-processing of the Smartphone Image Denoising Dataset (SIDD) (Abdelhamed, Lin & Brown, CVPR 2018), with ground truth re-estimated from the original DNG bursts and a new train/test split.

The dataset is published in two modalities, each self-contained with its own train and test data and its own README:

Raw/	Bayer-mosaic float32 arrays, black level subtracted and left unclipped — 23 GB
SRGB/	8-bit RGB PNGs rendered through the camera pipeline — 8.1 GB

Raw/                        SRGB/
  README.md                   README.md
  train/                      train/
    Scene_Instances.txt         Scene_Instances.txt
    Data/{scene}/*.MAT          Data/{scene}/*.PNG
  test/                       test/
    Benchmark*Blocks*.mat       Benchmark*Blocks*.mat
    BenchmarkBlocks32.mat       BenchmarkBlocks32.mat
    test_scenes_info.txt        test_scenes_info.txt

Each train/ mirrors the official SIDD_Medium_Raw / SIDD_Medium_Srgb layout, so existing SIDD loaders work with little change.

The train/test split

Of the 200 SIDD instances, 166 are still served by the official distribution. Samsung S6 captures whose RAW data was already clipped at black on the device are excluded, leaving 144 valid instances, split 115 train / 29 test (~20% test).

Test selection allowed at most one instance per (scene number, camera) pair, kept per-scene coverage proportional, and favoured ISO ≥ 400.

	train	test
Instances	115	29
Cameras	IP 49, GP 26, N6 17, S6 13, G4 10	GP 9, G4 7, N6 6, IP 6, S6 1
ISO	50 – 10,000	400 – 6,400
Illuminant	3200 K, 4400 K, 5500 K	3200 K, 4400 K, 5500 K
Brightness	normal 65, low 48, high 2	normal 17, low 12

Train ships as full-resolution images, 2 per instance (230 pairs). Test ships as 32 blocks of 256×256 per instance (928 blocks) — see either modality's README for how the blocks are extracted. The test instances' full-frame images are not published.

Instance names follow the SIDD convention:

0001_001_S6_00100_00060_3200_L
  |    |   |    |     |    |  |
  |    |   |    |     |    |  brightness code (L / N / H)
  |    |   |    |     |    illuminant temperature (K)
  |    |   |    |     shutter speed
  |    |   |    ISO
  |    |   camera
  |    scene number
  instance id

Source data

The raw DNG bursts that this ground truth was estimated from — 150 frames per instance — are in MhdMohammadi/SIDD-CC-DNG (train split).

Usage

# everything
hf download MhdMohammadi/SIDD-CC --repo-type dataset --local-dir SIDD-CC

# one modality
hf download MhdMohammadi/SIDD-CC --repo-type dataset --include "SRGB/*" --local-dir SIDD-CC

# just the evaluation blocks
hf download MhdMohammadi/SIDD-CC --repo-type dataset --include "*/test/*" --local-dir SIDD-CC

Licensing and attribution

Derived from the original SIDD release and subject to the SIDD terms of use. If you use this data, please cite both SIDD-CC and the original dataset:

@article{mohammadi2026low,
  title={Why Low-Light Cameras Go Color Blind: Removing Color Bias in Raw Denoising},
  author={Mohammadi, Mohammad and Honari, Sina and Tsogkas, Stavros and Aumentado-Armstrong, Tristan and Brown, Michael S and Mohomed, Iqbal and Derpanis, Konstantinos G and Levinshtein, Alex and Gilitschenski, Igor},
  journal={arXiv preprint arXiv:2607.11090},
  year={2026}
}

@inproceedings{abdelhamed2018high,
  title     = {A High-Quality Denoising Dataset for Smartphone Cameras},
  author    = {Abdelhamed, Abdelrahman and Lin, Stephen and Brown, Michael S.},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2018}
}