Color constancy measure

Let o be the constancy measure and let a2 be the color space measure. We can combine both measures into a color constancy measure m and compute the fraction... 

The results for this color constancy measure are shown in Table 13.11 and Table 13.12. The largest value for each set is printed in bold face. 

Table 13.11 Combined color constancy measure for image sets 1 through 4.

Figure 13.7 Color constancy measure as a function of the performance of the algorithms. The data for sets 1 through 4 are shown in graph (a). The data for sets 6 through 9 and sets 10 through 13 are shown in the two graphs (b) and (c).

Macbeth 5000 K fluorescent, a Philips Ultralume fluorescent, and a Sylvania Cool White fluorescent tube. Some color constancy algorithms try to find out what type of illuminant produced the particular color sensation, which was measured by the sensor. If we measure the entire power spectrum for a particular patch of our object and also know the type of illuminant used, then it is easy to compute the BRDF. Let L(X) be the measured power spectrum and let E A.) be the power spectrum of the illuminant. Assuming a Lambertian surface, where the BRDF is independent of the normal vector of the patch Nobj, and also independent of the normal vector that points to the direction of the light source Ni, we... 

Finlayson and Hordley (2001 a,b) have developed a method for color constancy at a pixel. They make two assumptions. One that the camera s sensors are sufficiently narrow band and that the illuminant can be approximated by a black-body radiator. If the sensor s response functions are not sufficiently narrow band, a sharpening technique can be used (Barnard et al. 2001 Finlayson and Funt 1996 Finlayson et al. 1994a,b). The power spectrum of a black-body radiator was described in Section 3.5. The radiance of a black-body radiator at a temperature T, measured in Kelvin, at wavelength /. is given by (Haken and Wolf 1990 ... 

Hurlbert and Poggio (1987, 1988) have shown that color constancy may be learnt from examples. Owing to computational constraints, they worked on individual lines of simulated color Mondrians. The color Mondrians were created by artificially illuminating a reflectance image of a color Mondrian with illumination gradients. In their model, the measurement made by the sensor / is proportional to the product of the illumination intensity L and the surface reflectance R. 

We will be using color-based object recognition to evaluate the quality of the color constancy algorithms. Where ground truth data is available, we can also compare the output of the different algorithms to the measured ground truth data. 

These are the two colors measured by the image sensor. In practice, intermediate colors also occur at the boundary of the patch however, this shall not be our concern here. This input stimulus is processed by a color constancy algorithm. In the following text, we will calculate the output color for each algorithm in turn. 

---