Abstract

The last decade has seen an astronomical shift from imaging with DSLR and point-and-shoot cameras to imaging with smartphone cameras. Due to the small aperture and sensor size, smartphone images have notably more noise than their DSLR counterparts. While denoising for smartphone images is an active research area, the research community currently lacks a denoising image dataset representative of real noisy images from smartphone cameras with high-quality ground truth. We address this issue in this paper with the following contributions. We propose a systematic procedure for estimating ground truth for noisy images that can be used to benchmark denoising performance for smartphone cameras. Using this procedure, we have captured a dataset, the Smartphone Image Denoising Dataset (SIDD), of ~30,000 noisy images from 10 scenes under different lighting conditions using five representative smartphone cameras and generated their ground truth images. We used this dataset to benchmark a number of denoising algorithms. We show that CNN-based methods perform better when trained on our high-quality dataset than when trained using alternative strategies, such as low-ISO images used as a proxy for ground truth data.

Papers

Abdelrahman Abdelhamed, Lin S., Brown M. S. "A High-Quality Denoising Dataset for Smartphone Cameras", IEEE Computer Vision and Pattern Recognition (CVPR), June 2018.

[PDF] [Bibtex]

@InProceedings{SIDD_2018_CVPR,
					author = {Abdelhamed, Abdelrahman and Lin, Stephen and Brown, Michael S.},
					title = {A High-Quality Denoising Dataset for Smartphone Cameras},
					booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
					month = {June},
					year = {2018}
				}

Abdelrahman Abdelhamed, Timofte R., Brown M. S., et al. "NTIRE 2019 Challenge on Real Image Denoising: Methods and Results", IEEE Computer Vision and Pattern Recognition Workshops (CVPRW), June 2019.

[PDF] [Bibtex]

@InProceedings{Abdelhamed_2019_CVPR_Workshops,
					author = {Abdelhamed, Abdelrahman and Timofte, Radu and Brown, Michael S. and others},
					title = {NTIRE 2019 Challenge on Real Image Denoising: Methods and Results},
					booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
					month = {June},
					year = {2019}
				}

Code

Ground-truth image estimation

A simple camera pipeline for rendering raw-RGB images into sRGB.

License

The dataset and the associated code repositories are under the MIT License.

Contact

For any questions, remarks, or comments, please contact: Abdelrahman Abdelhamed.

SIDD-Small Dataset

Download

Raw-RGB images only (~10 GB) [ Mirror 1 | Mirror 2 ]
MD5: c033f580fb64aa549679d056d1cf796a SHA1: 932c39f4fc5410dda934a3493ddb46c31b8a5e90

sRGB images only (~6 GB) [ Mirror 1 | Mirror 2 ]
MD5: 796971867583bf14677dcae510e52538 SHA1: 5a4aa6aa7abcf7b0b56c88ad578fea9a4ff77935

We provide a small version of the dataset that consists of 160 image pairs (noisy and ground-truth) representing 160 scene instances. These images can be used for training/learning purposes.

For each image, the following is provided:

Noisy Raw-RGB image (.MAT).

Black Level subtracted, normalized to [0, 1].

Ground truth Raw-RGB image (.MAT).

Black Level subtracted, normalized to [0, 1].

Noisy sRGB image (.PNG).

Gamma corrected, without any tone mapping.

Ground truth sRGB image (.PNG).

Gamma corrected, without any tone mapping.

Metadata extracted from the DNG file (.MAT).

For example, black and saturation levels, as-shot neutral, noise level function, etc.

Camera Pipeline (Rendering from raw-RGB to sRGB)

We provide a simple and light-weight camera image processing pipeline implemented in MATLAB. This pipeline can be used to render the raw-RGB images into sRGB images with optional tone mapping.

[ GitHub | MathWorks ]

SIDD-Medium Dataset

The SIDD Medium dataset is similar to SIDD Small dataset excpet that SIDD Medium consists of 320 image pairs (noisy and ground-truth), two image pairs from each scene instance.

Download

SIDD-Full Dataset

We provide 80% (~24,000 images) of the dataset for training/learning purposes. The rest of the dataset is held for the benchmark.

Below, we provide links to 160 scene instances. For each scene instance, the following is provided:

150 noisy Raw-RGB image (.MAT).

Black Level subtracted, normalized to [0, 1].

150 ground truth Raw-RGB image (.MAT).

Black Level subtracted, normalized to [0, 1].

150 noisy sRGB image (.PNG).

Gamma corrected, without any tone mapping.

150 ground truth sRGB image (.PNG).

Gamma corrected, without any tone mapping.

150 metadata extracted from the DNG file (.MAT).

For example, black and saturation levels, as-shot neutral, noise level function, etc.

Note: scene instances appearing in gray are held for benchmark.

Image directory naming convention

Each image is stored in a directory with the name of the scene instance as follows:
[scene-instance-number]_[scene_number]_[smartphone-camera-code]_[ISO-level]_[shutter-speed]_[illuminant-temperature]_[illuminant-brightness-code]
where "smartphone-camera-code" is one of the following:

GP: Google Pixel
IP: iPhone 7
S6: Samsung Galaxy S6 Edge
N6: Motorola Nexus 6
G4: LG G4

and "illuminant-brightness-code" is one of the following:

L: low light
N: normal brightness
H: high exposure

For example, the following directory name:
0052_002_S6_01600_01000_5500_N
means that this is scene instance 52, from scene 2, captured be Samsung Galaxy S6 Edge (S6), using ISO level of 1600, using shutter speed of 1000 (i.e., exposure time of 1/1000 second), under illuminant temperature of 5500K, under normal (N) brightness.

Useful metadata

Bayer pattern for each camera: bayer_patterns.csv

Camera noise level functions (NLF) as extracted from DNG files: noise_level_functions.csv

DNG images

DNG images corresponding to noisy raw-RGB images

The first four digits in each DNG archive should match the corresponding scene instance number

For example: "0001_DNG.zip" corresponds to "0001_001_S6_00100_00060_3200_L"

Download

Text files containing all links to the files below that can be used for batch-downloading (e.g., using custom scripts or download managers):
Mirror 1 | Mirror 2

Download individual scene instances

SIDD Benchmark

A new version of SIDD Benchmark (SIDD+) was hosted as a challenge at the New Trends in Image Restoration and Enhancement (NTIRE 2020) workshop in conjunction with CVPR 2020.
The participating solutions and results will be published in the challenge report in the CVPR 2020 Workshop proceedings.

Challenges can be accessed at the following Codalab competitions:
NTIRE 2020 Real Image Denoising Challenge - Track 1 - rawRGB
NTIRE 2020 Real Image Denoising Challenge - Track 2 - sRGB

Download

SIDD Benchmark Code v1.2 (9 KB) [ Mirror 1 | Mirror 2 | Mirror 3 | Mirror 4 ]

SIDD Benchmark Data as single .mat arrays of dimensoins [#images, #blocks, height, width, #channels]:
Noisy raw-RGB data: [ Mirror 1 | Mirror 2 ]
Noisy sRGB data: [ Mirror 1 | Mirror 2 ]

SIDD Validation Data and Ground Truth as single .mat arrays of dimensoins [#images, #blocks, height, width, #channels]:
Noisy raw-RGB data: [ Mirror 1 | Mirror 2 ]
Noisy sRGB data: [ Mirror 1 | Mirror 2 ]
Ground-truth raw-RGB data: [ Mirror 1 | Mirror 2 ]
Ground-truth sRGB data: [ Mirror 1 | Mirror 2 ]

SIDD Benchmark Data (full-frame images, 1.84 GB) [ Mirror 1 | Mirror 2 | Mirror 3 | Mirror 4 ]
MD5: decd113eaf99a8dbd1dbb7f7c9dafedd SHA1: b8092d990139f41b6da97b4afa679a2876de53bd

Description

The SIDD Benchmark consists of 40 images representing 40 scene instances. These images can be used to benchmark denoising methods.

For each image, the following is provided in one directory:

Noisy Raw-RGB image (.MAT). Black Level subtracted, normalized to [0, 1].
Noisy sRGB image (.PNG). Gamma corrected, without any tone mapping.
Metadata extracted from the DNG file (.MAT). For example, black and saturation levels, as-shot neutral, noise level function, etc.

The PSNR and SSIM values are calculated only on 32 blocks of size 256 by 256 pixels. The block positions are provided in a file named "BenchmarkBlocks32.mat".

Follow the instructions in the Code_v/_ReadMe.txt file to

demonstrate the code,
run your own denoising method on the benchmark, and/or
submit your results for evaluation.

Upload your results

Click here to submit your results

Benchmark Results

Results of the NTIRE 2019 Challenge on Real Image Denoising can be found in this paper.

The following tables show the benchmark results published in the paper.

Results of denoising in raw-RGB space

Raw-RGB results

Results of denoising in sRGB space

sRGB results