GPU Color Constancy

A sensor located inside a digital camera is only able to measure the light which is reflected by an object. The reflected light varies with the spectral power distribution of the illuminant. Hence, images taken with a digital camera may show a strong color cast if an incorrect white balance setting has been chosen. Such a color cast may also be due to an automatic white balance not working correctly. In contrast, colors perceived by a human observer appear to be approximately constant. Algorithms for automatic white balance try to mimic this ability and compute a color corrected image which appears to have been taken under an illuminant with a uniform power distribution. I show how color constancy algorithms can be implemented very efficiently on modern graphics processing units. A sensor inside a digital camera measures the light which is reflected from the objects of the scene. Some of the light is absorbed, the remainder is reflected and is able to enter the lens of the camera where it is measured. The reflected light varies with the type of illuminant used. Some light sources emit more light towards the red and green part of the spectrum and therefore appear to be very yellowish. If such an illuminant is used, then the scene will come out very yellowish in a photograph. The color temperature of daylight also varies during the day. Digital cameras can use post-processing to remove such a color cast. They try to compute an image which appears to have been taken under an illuminant with a uniform power spectrum. In digital photography, this is termed automatic white balance. Many digital cameras also allow the user to select a particular color temperature. Some have pre-settings for sunlight, cloudy sky, neon light or illumination by a light bulb. In contrast to a machine sensor, the colors perceived by a human observer appear to be remarkably constant. This ability is called color constancy [18, 5]. For machine vision applications it is very important to mimic this ability. For instance, color based object recognition becomes very difficult if objects appear to change their color based on the type of illuminant used. Therefore, it makes sense to first compute a color constant descriptor [12]. A number of algorithms have been developed to address the problem of color constancy. The problem can only be solved if some assumptions are made. I present a color constancy method based on local space average color which is easily implemented on a graphics processing unit (GPU). Algorithms based on local space average color have been shown to work very effectively in object recognition tasks [6]. The advantages of this method are: (1) it also works in the presence of multiple spatially varying illuminants and (2) it is readily implemented with 3 or 4 lines of code. It can be used to (1) remove a color cast from textures provided that the assumptions made by the algorithm are fulfilled and (2) provide color constancy for GPU accelerated computer vision, robotics, animation or interactive gaming applications. I also show how intrinsic images, which only depend on the reflectance component, are efficiently computed through a pixel shader. The algorithm assumes that the illuminant can be approximated to have the spectral power distribution of a black-body radiator. Such intrinsic images are free from shadows which make them very useful for image segmentation.