The Rehabilitation of MaxRGB

The poor performance of the MaxRGB illuminationestimation method is often used in the literature as a foil when promoting some new illumination-estimation method. However, the results presented here show that in fact MaxRGB works surprisingly well when tested on a new dataset of 105 high dynamic range images, and also better than previously reported when some simple pre-processing is applied to the images of the standard 321 image set [1]. The HDR images in the dataset for color constancy research were constructed in the standard way from multiple exposures of the same scene. The color of the scene illumination was determined by photographing an extra HDR image of the scene with 4 Gretag Macbeth mini Colorcheckers at 45 degrees relative to one another placed in it. With preprocessing, MaxRGB’s performance is statistically equivalent to that of Color by Correlation [2] and statistically superior to that of the Greyedge [3] algorithm on the 321 set (null hypothesis rejected at the 5% significance level). It also performs as well as Greyedge on the HDR set. These results demonstrate that MaxRGB is far more effective than it has been reputed to be so long as it is applied to image data that encodes the full dynamic range of the original scene. Introduction MaxRGB is an extremely simple method of estimating the chromaticity of the scene illumination for color constancy and automatic white balancing based on the assumption that the triple of maxima obtained independently from each of the three color channels represents the color of the illumination. It is often used as a foil to demonstrate how much better some newly proposed algorithm performs in comparison. However, is its performance really as bad as it has been reported [1,3-5] to be? Is it really any worse than the algorithms to which it is compared?1 The prevailing belief in the field about the inadequacy of MaxRGB is reflected in the following two quotations from two different anonymous reviewers criticizing a manuscript describing a different illumination-estimation proposal: “Almost no-one uses Max RGB in the field (or in commercial cameras). That this, rejected method, gives better performance than the (proposed) method is grounds alone for rejection.” “The first and foremost thing that attracts attention is the remarkable performance of the Scale-by-Max (i.e. White-Patch) algorithm. This algorithm has the highest performance on two of the three data sets, which is quite remarkable by itself.”   Paper’s title inspired by Charles Poynton, “The Rehabilitation of Gamma,” Proc. of Human Vision and Electronic Imaging III SPIE 3299, 232-249, 1998. We hypothesize that there are two reasons why the effectiveness of MaxRGB may have been underestimated. One is that it is important not to apply MaxRGB naively as the simple maximum of each channel, but rather it is necessary to preprocess the image data somewhat before calculating the maximum, otherwise a single bad pixel or spurious noise will lead to the maximum being incorrect. The second is that MaxRGB generally has been applied to 8-bit-per-channel, non-linear images, for which there is both significant tone-curve compression and clipping of high intensity values. To test the pre-processing hypothesis, the effects of preprocessing by median filtering, and resizing by bilinear filtering, are compared to that of the common pre-processing, which simply discards pixels for which at least one channel is maximal (i.e., for n-bit images when R=2n-1 or G=2n-1 or B=2n-1). To test the dynamic-range hypothesis, a new HDR dataset for color constancy research has been constructed which consists of images of 105 scenes. For each scene there are HDR2 (high dynamic range) images with and without Macbeth mini Colorchecker charts, from which the chromaticity of the scene illumination is measured. This data set is now available on-line. MaxRGB is a special and extremely limited case of Retinex [6]. In particular, it corresponds to McCann99 Retinex [7] when the number of iterations is infinite, or to path-based Retinex [8] without thresholding but with infinite paths. Retinex and MaxRGB both depend on the assumption that either there is a white surface in the scene, or there are three separate surfaces reflecting maximally in the R, G and B sensitivity ranges. In practice, most digital still cameras are incapable of capturing the full dynamic range of a scene and use exposures and tone reproduction curves that clip or compress high digital counts. As a result, the maximum R, G and B digital counts from an image generally do not faithfully represent the corresponding maximum scene radiances. Barnard et al. [9] present some tests using artificial clipping of images that show the effect that lack of dynamic range can have on various illumination-estimation algorithms. To determine whether or not MaxRGB is really as poor as it is report to be in comparison to other illumination-estimation algorithms, we compare the performance of several algorithms on the new image database. We also find that two simple preprocessing strategies lead to significant performance improvement in the case of MaxRGB. Tests described below show that MaxRGB performs as well on this new HDR data set as other representative and recently published algorithms. We also find that two simple pre-processing strategies lead to significant performance improvement. The results reported here extend those of an earlier study [10] in a number of ways: the size of the dataset   2 Note that the scenes were not necessarily of high dynamic range. The term HDR is used here to mean simply that that full dynamic range of the scene is captured within the image. 3 www.cs.sfu.ca/~colour/data  Page 1 of 4