Illuminant Estimation for Color Constancy:

Abstract:

Color constancy is a well studied topic in color vision. Methods are generally categorized as: 1) low-level statistical methods; 2) gamut-based methods; 3) and learning-based methods. In this work, we distinguish methods depending on whether they work directly from color values (i.e. color domain) or from values obtained from the image spatial information (e.g. image gradients/frequencies). We show that spatial information does not provide any additional information that cannot be obtained directly from the color distributions and that the indirect aim of spatial domain methods is to obtain large color differences for estimating the illumination direction. This finding allowed us to develop a simple and efficient illumination estimation method that chooses bright and dark pixels using a projection distance in the color distribution and then applies PCA to estimate the illumination direction. Our method gives state-of-the-art results on existing public illumination datasets as well as on our newly collected dataset containing 1736 images from 8 different cameras.

Results on existing Datasets:

Note:

Newly Collected Datasets:

For our own collected image datasets from 9 different cameras, some example scenes of the datasets are shown below.

The data is available in three formats: (1) camera-specific RAW format free of any manipulation, (2) DCRAW-converted linear images which can be read and used directly, as well as (3) camera JPEG outputs for the purpose of viewing the image content. In order to get original linear raw data, DCRAW command line arguments '-D -4 -T' was used. No further demosaicing was done, but only downsampling according to camera's bayer pattern to create a color image. To be noted, care had been taken for the Fujifilm X-M1 camera, which has a novel 6*6 bayer pattern. For each image, co-ordinates for the color-checker inside the image are also provided, as well as small region masks for every color patch. The ground truth illuminants were obtained from the difference of the two most brightest achromatic patches containing no saturated value to get rid of the effects from the dark level and saturation level of the sensor. In order to use those images, dark level has to be subtracted from the original images, which is provided in the ground truth .mat file.

Datesets and results are divided according to the cameras: (click on the camera name to see the according results)

And another extra camera with smaller number of images is also provided here:

Results on our newly collected Datasets (click on "proposed" to download our results):

Timing Comparison:

Source Code:

Acknowledgements:

This study was funded by Agency for Science, Technology and Research (grant No. 1121202020). We thank our photographer Mr. Looi Wenhe (Russell) for his hard work and effort in helping collect our dataset.