s24-raw-srgb-dataset

🏞️ S24 Raw-sRGB Dataset

Figure: The S24 Raw-sRGB Dataset contains 3,224 raw images, each with a denoised version manually adjusted for accuracy. Each raw image is paired with an sRGB image from the in-camera ISP (Pro mode) and an expert-rendered sRGB image from Adobe Lightroom that incorporates local tone mapping for aesthetic enhancement. Additionally, five more sRGB images are provided using different Lightroom picture styles. Each sample includes capture-time metadata and both neutral and user-preference white balance ground-truth.

🧾 Overview

The S24 Raw-sRGB Dataset is a large-scale collection of paired raw and sRGB images captured using the Samsung Galaxy S24 Ultra’s main camera. It contains 3,224 raw images, each paired with multiple sRGB renditions:

• In-camera ISP (Pro mode): sRGB images generated by the camera’s on-device ISP.
• Adobe Lightroom (manual editing): Expert-rendered sRGB images with manual adjustments, including local tone mapping, to enhance visual quality.
• Adobe Lightroom presets: Each image is also rendered using five different Lightroom picture style presets.

Figure: Our dataset includes sRGB images produced by the in-camera ISP and expert-rendered sRGB images from Adobe Lightroom, with local tone mapping applied for enhanced aesthetics..

Figure: In addition to the expert rendering (style #0), our dataset includes sRGB images with multiple styles, which can serve as ground truth for picture style transfer or raw-to-multi-style sRGB rendering.

📌 Neural-ISP Results: For raw-to-sRGB benchmarking on both the default style (manual expert editing) and the other picture styles, refer to Tables 10 and 11 in our paper.

The dataset also includes:

• Denoised raw images generated using Adobe Lightroom’s AI denoiser.

• Ground-truth white balance annotations, including both neutral and user-preference variants.

• Binary masks identifying regions affected by secondary light sources.

• Scene and light classes, specifying the type of scene (e.g., indoor, outdoor) and the lighting condition (e.g., natural, artificial).

Figure: Our dataset includes diverse scenes captured under various weather conditions (sunny, cloudy, rainy, and snowy) and lighting conditions (indoor, daylight, sunset/sunrise, and night). For each example, we show raw images (gamma-corrected for better visualization) alongside their sRGB counterparts.

🗃️ Data Structure

The dataset is organized into the following folders:

• train: Training set
• test: Testing set
• val: Validation set

Within each set, the following subfolders are included:

• raw_images: 16-bit PNG files of raw images (3 channels) after demosaicing and black-level normalization.

• denoised_raw_images: 16-bit PNG files of raw images after denoising, generated using Adobe Lightroom's AI denoiser (with manual adjustments for accuracy).

• srgb_images: 8-bit JPG files of camera-generated sRGB output images (captured using Pro mode on S24 Ultra with the in-camera ISP).
  Note: A slight pixel shift exists between the sRGB and raw images due to internal cropping applied by the camera’s ISP.

• srgb_images_style_0: 8-bit JPG files of manually rendered sRGB images using Adobe Lightroom, with local tone mapping adjustments.

• srgb_images_style_X: 8-bit JPG files of sRGB images rendered using five different Adobe Lightroom presets (where X is 1, 2, ..., 5).

• masks: 8-bit PNG masks identifying regions lit by a different illuminant than the dominant one used to obtain the white-balance ground-truth.

• blur_masks: 8-bit binary masks indicating regions where personal information has been blurred (1 = blurred, 0 = not blurred).

• data: JSON files containing associated capture information, including ISO, shutter speed, time of day (computed from EXIF location and time), orientation, and more:

  • gt_illum: Neutral white-balance ground-truth.

  • pref_illum: User-preference adjusted ground-truth.

  • scene_class: Manually annotated scene types (e.g., daylight, sunset/sunrise, indoor, night).

  • light_class: Type of light (natural or artificial).

  • ccm: Color correction matrix (CCM) to convert white-balanced raw colors to linear sRGB, extracted from the DNG metadata.

  • cam_illum and cam_daylight_illum: Camera-estimated scene illuminant (RGB) and the camera’s calibrated daylight illuminant (RGB), respectively.

  • capture_metadata: Includes capture-related and time-based features (see Section 2.1 in our paper). Specifically:

    • iso, exposure_time, exposure_bias, shutter_speed, flash
    • Time-related features:
      • prob_X and is_before_X, where X ∈ {sunrise, sunset, dusk, dawn, noon, midnight}
      • prob_X represents the probability that the capture occurred around the solar event X
      • is_before_X indicates whether the capture occurred before or after the event X

  • noise_stats: Noise statistics as explained in Section 2.1 of the paper. Includes:

    • rgb_mean, rgb_std: Mean and standard deviation of RGB channels from the absolute difference between denoised and noisy raw images
    • mean, std: Mean and standard deviation of the weighted average RGB channels

  • snr_stats: SNR statistics as described in Section 2.1. Includes:

    • mean, std: Mean and standard deviation of the signal-to-noise ratio image

  • additional_metadata: Additional metadata extracted from the DNG file, such as camera make and model, pre-calibrated color space transformation matrices, etc.

📝 Notes

• All personal information in the images (both raw and sRGB) has been blurred out to ensure privacy.
• The sRGB images rendered using Adobe Lightroom were based on the camera’s illuminant vector.

📄 Citation

If you use this code or dataset in your research, please cite our paper:

@inproceedings{afifi2025time-aware-awb,
  title={Time-Aware Auto White Balance in Mobile Photography},
  author={Afifi, Mahmoud and Zhao, Luxi and Punnappurath, Abhijith and Abdelsalam, Mohamed Ashraf and Zhang, Ran and Brown, Michael S.},
  booktitle={The IEEE International Conference on Computer Vision (ICCV)},
  year={2025}
}

🔗 Dataset Access

• Training set

• Validation set

• Testing set

📬 Contact

For questions or inquiries, please contact:

• Mahmoud Afifi (m.afifi1@samsung.com, m.3afifi@gmail.com)
• Luxi Zhao (lucy.zhao@samsung.com, lucyzhao.zlx@gmail.com)