General purpose function for creating an object-based scene

[wandell]

Suppose we have a monochrome image of a car on a neutral background. We would like to make an image of that car at different contrast values, something like

 scene =  Background + contrast*CarImage

Suppose the background is uniform = 1, for the moment. Then we could adjust the carImage to zero mean.

  carImage = carImage - mean(carImage(:));

And scale the carImage so that its smallest (most negative) number is greater equal -1.

  s = min(carImage(:)); 
  scaledCarImage = (carImage/abs(s));
  scene = Background + contrast*scaledCarImage

where 0 < contrast < 1. In this case, when contrast = 1, the scene will vary between [0,X] and the mean will be 1. In this way, when we scale by P to set the number of photons (or energy) by multiplying, we always have positive photons (or energy) and the mean is equal to the scale factor.

In the future, we might imagine that Background is an arbitrary texture pattern. And then we would have to adjust how we choose the scale factor, s, to guarantee we get a physically realizable image.

The other issue is how are doing this with color.

[DavidBrainard]

We have had to deal with this, or at least a very similar, issue when we want to rapidly vary the properties of one object in a scene, as for example when a subject adjusts the color of one object in a scene to match another. The way we have approached this is to render a number of basis images, which differ only in the reflectance of one object. So, for example, if you have two versions of the scene with the car, where one has a low reflectance and the other has a high reflectance. Call the first image I1 and the second I2. Then you can construct k*I1 + (1-k)*I2. As you sweep k between 0 and 1, you sweep the reflectance of the car between the high and low reflectance.

The same principle applies to varying color. In principle, three basis images is enough to produce anything inside their gamut; in practice we've found it's good to use more to reduce artifacts that arise from some sort of numerical issues. We settled on 8 the last time we did this.

Nicolas put together a nice package that manages this sort of mixing.

The underlying idea is due to a paper by a guy named Griffin in 1999. There is a story about that paper that involves Bosco Tjan that I will tell you sometime. Our version is described in the methods section
of the two attached papers.

Xiao_Brainard_08.pdf
i1534-7362-15-13-3-2.pdf