Color Datasets 

Middlebury Color Datasets

  1. Registered Color Checker images
  2. We provide registered views of two color calibration targets, X-Rite's 24-patch ColorChecker and 140-patch Digital ColorChecker SG. Each pattern was photographed:
    • with 35 different camera models, 24 of which support both RAW and JPEG, while 11 support only JPEG;
    • with fixed "Tungsten" white balance (wb1); for a subset of cameras also with "auto" white balance (wb2);
    • under two fixed lighting conditions: using 3200K Tungsten (i1) and 4800K Daylight (i2) photo flood light bulbs;
    • with 4 different exposures: stops -2 (e1), -1 (e2), 0 (e3), and +1 (e4).
    The RAW images were rendered as PNGs in standard linear RGB colorspace with the program dcraw, using the camera's white balance multipliers. All images were aligned using homographies based on the corner coordinates. We provide both point-sampled and "smoothed" (4x linearly down-sampled) versions of these images. The former represent true samples of the original intensities, sensor noise, and JPEG compression artifacts, while the latter attenuate such effects. The final registered images were created by cropping and compositing the individual color squares. We also include noise estimates for each image and color channel (the average standard deviation within each constant patch).

    RAW + JPEG
    24 camera models
    JPEG only
    11 camera models
    Sampled (thumbnails) (123M, contents) (28M, contents)
    Smoothed (thumbnails) (41M, contents) (11M, contents)
    Sampled (thumbnails) (76M, contents) (17M, contents)
    Smoothed (thumbnails) (26M, contents) (6M, contents)


  3. Registered RAW-JPEG pairs of natural scenes
  4. We provide 85 registered RAW/JPG pairs of natural scenes taken with 12 different camera models; both "scaled" (linearly down-sampled) versions attenuating noise and compression effects, and "cut" versions composited from patches of the original images.

    Scaled (thumbnails) (139M, contents)
    Cut (thumbnails) (136M, contents)


  5. Registered JPEG images of planar surfaces
  6. We provide registered images of planar surfaces taken with 5 different camera models, from different viewpoints, under different lighting, and with different white balance and exposure. The views were registered using Chuck Stewart's GDB-ICP software after correcting for radial lens distortion.

    Link to Datasets
    (zip files available upon request)


How to cite our datasets:

We grant permission to use and publish all images on this website. If you do, we request that you cite the following paper:

A. Chakrabarti, D. Scharstein, and T. Zickler.
An empirical camera model for Internet color vision.
In British Machine Vision Conference (BMVC 2009), London, UK, September 2009.


Many thanks to Porter Westling for his help in generating datasets #1 and #2, and to Scott Wehrwein and Kelvin Gorekore for their help in generating datasets #3.











Last modified: June 23 2020 by Daniel Scharstein