Verification and extension of a camera-based end-user calibration method for projection displays

We evaluate, analyse and propose improvements to a previously published end-user calibration method for projectio n devices (Bala and Braun, CIC 2006). We focus on the estimation of the displays tone response curve, using only an uncalibra ted consumercamera. The results show that the method is accurat e, depending on both the projector and the camera used. We found that the method is accurate enough for most end-user applica tions. A weakness of this method is the wrong estimation of th e projectors black level, which significantly affects the est ima ion of the camera response curve. Introduction Nowadays, projection displays are widely used but rarely calibrated properly [1]. The resulting lack of color accura cy often leads to not only a loss of visual appeal in the presente d material, but also in many cases to a loss of intended meaning . Accurate colorimetric calibration of projection displays is technically challenging and requires expensive specialist equ ipment. Bala and Braun [2, 3] recently presented an alternative endu ser method for projection display calibration. The work presen ted in this paper aims is to perform an independent evaluation and a nalysis of this method, and to propose and evaluate improvement s. Proposed methodology The Bala method [2, 3] aims to achieve a good calibration for projection displays using no equipment other than an unc alibrated consumer digital camera that anyone could have at hom e. This removes the need for any radiometric or colorimetric me asurements in the calibration process. It can be assumed that the system primaries are sRGB [5]. Their proposal is to calibrate the camera relatively to the display, using the following information: (a) The 50% luminance point of the projector [2, 3, 4], estimated visually using a matching pattern. (b) Considering the maximum luminance at 1 and the true black at 0. (c) Considering the black level of the projector at 2% of the maximum luminance [2, 3] (for a dim surrounding). This information is used to estimate the camera response curve, using spline interpolation. Based on this estimatio n, the camera can replace a more sophisticated photometric measur ement device and can be used to estimate the projectors tone re sponse curve, by taking a picture of a pattern displayed on th e screen. We invite the reader to read [2, 3] in order to get all d etails. In the present work, the Bala method is re-implemented with the aim of identifying strengths and weaknesses of the a pproach. Moreover, we propose to enhance the original method as 1E.B.M. is currently working at 3D-Perception; Asker, Norwa y. 2J.-B.T. is also researcher at the Le2i, University of Bourgo gne; Dijon, France. follows: (a) The original method assumes that the normalized grey level response curve is the same as the normalized curves for each primary. We propose to separate estimations of the thre primary color channels response curves through duplicatio n of the calibration procedure per channel. (b) Another improvement is to increase the number of visually determined luminance levels from one to three by addi ng targets for 25% and 75% luminance to the original 50% level, obtained by a haltoning like process, similarly to the full v isual calibration proposed in [6]. Experimental setup and results The Bala method and its proposed enhancements were implemented and tested with two different digital cameras: th e Nikon D200 DSLR and the Fujifilm Finepix S7000 compact type camera, and two different projectors with different techno logies: a Panasonic AX-PT100E LCD and a Projectiondesign Action one DLP. All experiments were done with default hardware set tings and no gamma correction performed by image source computer. Figure 1. Camera response estimation using different black levels. Giving a wrong black point makes the estimation to vary strongly for the luminance below the 50% of the projector. Throughout our experimentation with the Bala method [7] it became apparent that the method can be accurate enough for some applications, but performances are largely dependant on three main factors: (a) The estimation of the cameras response curve is based on four points. Two of these are the absolute black point and t he projectors black point which are close together. The proxim ity Table 1: Average ∆L between the real projector response curve and the estimated one (the 256 possible values were measured and estimated), depending on the method used. Method Projector ∆L ∆L Red ∆L Green ∆L Blue RGB average ∆L Original LCD 3.47 3 matched luminances LCD 2.14 Original DLP 1.64 3 matched luminances DLP 0.59 Separate Channel match w/ 1 lum. match LCD 1.83 3.03 2.51 2.46 Separate Channel match w/ 3 lum. match LCD 1.48 2.30 1.92 1.90 Separate Channel match w/ 1 lum. match DLP 1.90 1.05 2.96 1.97 Separate Channel match w/ 3 lum. match DLP 1.89 0.96 2.01 1.62 of these points strongly influences the shape of the interpol ation function. We have concluded that the projectors black level ne ds to be measured or estimated with good accuracy when using thi s method (see Figure 1). We propose no solution for an estimation of this parameter without an accurate measurement devi ce. This is a major weakness in the approach as it moves away from the initial chain of thought on replacing expensive color an d luminance measurement equipment with the digital camera. We do not put aside the fact that for other interpolation method s it could work well with a generic black level. However, we wonde r how increasing the number of visual pattern (in low luminanc es) while simply removing the projector black point to estimate th camera response curve can achieve good result. (b) Secondly , the observers precision in the visual matching task will determ ine a third point in the interpolation of the camera response cur ve. This will have significant influence on the estimation of came ra tone response. This is the weakness of every visual calibrat ion method. Note that the visual estimation of the 50% luminance for the blue channel is a harder task for the human visual system compared to the higher wavelengths of red and green. Furthermore, a ”grey balancing” [8] method can not be used as the projectors are used to show a large chromaticity shift with t he variation of input for the pure primaries. (c) The third factor is the cameras capability to differenti ate between projected luminances in the calibration patter n. If it lacks accuracy when recreating on-screen luminance differ ences it will not give the information needed to estimate projecto r tone response. When testing the method it became clear that the Fu jiFilm camera had severe problems with capturing suitable ima ges for use with this method. The captured images seemed to be either saturated in brighter areas or the darker patches were i ndistinguishable from each other, and the resulting estimated c urves were not good. Major efforts were put into experimentation w ith camera settings without achieving better images. As a conse quence, the camera was not found suitable to be used with this method; hence no further results from this camera will be reported. Although such critical factors were identified, the method shows good results (Table 1). These results are presented fo r a dark surrounding. The∆L is computed from the measured response curve and the estimated one, for a full ramp (256 value s) of grey level patches or for each independent channel. Note t hat Bala et al worked in dim surrounding. In such a case, it is possible that the estimation of 2% luminance for the black level is better. To be fair in our comparison, and to present comparab le results, all the methods presented used the same level of bla ck, which is the black level of the projector measured with a spec troradiometer Minolta CS-1000. The ∆L has been computed for two luminancesL1 andL2 as∆L = √ (L1−L2). Figure 2 shows the estimated tone reproduction curve for the Projectiondesign DLP projector using the original meth od. Here an average ∆L luminance difference of 1.64 from the projectors measured response was achieved. Figure 3 shows resu lt when using the extended method with three visually matched l uminance levels instead of one. This shows an even closer matc h to the measured response with an averaged ∆L difference of only 0.59. With the LCD projection device, we obtained 3.47 ∆L for the original method to 1.90 using both improvements (calibration of each primary, with 3 visual patches). The Fi gures Figure 2. Normalized luminance grey level response curve estimation for the original method (plain line) versus the measured one (dashed line) for the DLP projector, function of the input digital value. Figure 3. Normalized luminance grey level response curve estimation for the three luminance matchs method (plain line) versus the measured one (dashed line) for the DLP projector, function of the input digital value. 4 and 5 are showing the estimated tone response curve for the LCD for respectively the original and the extended method us ing three visually matched luminance levels. One can notice the same thing that for the previous display: we reduced the aver age error to 2.14. It appears that the independent estimation of the blue channel response curve for the both projectors show a ∆L of 2.52 and 2.96 for one luminance match and of 1.92 and 2.01 for three luminance match. It is supposed to be the worth case as the visual system is not good to distinguish luminance changing in the short wavelength. Using three luminance matching points, w e improve the estimation of the blue channel response curve, w hile for the red channel, this does not change the result very much . Doing this for the DLP projector green channel, we do not improve the estimation quality. However, the LCD projector sh ows a large error of matching for this channel, and using three po ints is beneficial for its estimation. Note that this error in matc hing luminance for the green channel, over t