21 Nov 2003 raph   » (Master)

chema

It's hard to believe that Chema Celorio has indeed passed away. We spent quite a bit of time together at the first Linux Printing Summit. I was really impressed by his passion, and have enjoyed working with him since. He will be sorely missed.

Multidimensional interpolation

The Argyll color management system uses a very sophisticated technique based on thin-plate splines to for its profile creation. The core operation is to take the colors for a few hundred patches and fit it to a smooth multidimensional function, typically from R^4 to R^3 (CMYK tint values to a 3-dimensional colorspace such as Lab).

I'm seeing some wiggliness in the profiles I'm generating. I'm not sure why; it might be measurement error on my part. But I'm starting to suspect it may be the interpolation algorithm itself. I've read a recent survey, but it doesn't provide clear guidance on which technique provides the smoothest results.

My own personal inclination would be to create the interpolation function by solving the heat-diffusion equation with a source at patch point, so as to fit the measurements. I'm convinced that this approach would give an optimum level of smoothness, but I see problems at the boundaries, even when using the free boundary assumption.

To simplify the discussion, assume that we have to extrapolate beyond the range of colors measured. Even in the simple case of two points, one at (0, 0) and another at (1, 1), the only solution I'd consider optimally smooth is the straight line f(x)=x. Yet, a straightforward solution of the heat diffusion equation with sources at 0 and 1 is going to have a slope of zero at the asymptotes.

I'm thinking of two possibilities: first, there might be a different boundary assumption that converges to a slope of 1 at both asymptotes. It's not too hard to imagine that somebody else might have come across such a thing; you'd need to to minimize the global curvature (magnitude of second derivative) for the solution.

Another possibility is to solve the heat diffusion equation for the derivative of the desired solution. In the 1-D case, it's pretty easy to see how it would work. For your sources, you use a derivative-of-Gaussian shape. For the example above, the derivative would obviously converge to a constant 1 (the free boundary assumption for the derivative seems like the right thing here). However, it's not immediately obvious to me how to generalize this to higher dimensions (in fairness, rillian tried to explain the usual finite-difference multidimensional integration technique, but I didn't quite grasp it at the time; I'll ask again).

But it's possible I'm going over well-known territory, and a reader better versed in diffusion problems can just tell me the answer. I'll find out soon, and even if not I'll report again on what I find.

Politics

My new favorite blog is Talking Points Memo. That's where I found a link to this video clip in which Wesley Clark kicks the ass of some lame Fox News reporter. One of the most satisfying bits of video I've seen in a while.

Latest blog entries     Older blog entries

New Advogato Features

New HTML Parser: The long-awaited libxml2 based HTML parser code is live. It needs further work but already handles most markup better than the original parser.

Keep up with the latest Advogato features by reading the Advogato status blog.

If you're a C programmer with some spare time, take a look at the mod_virgule project page and help us with one of the tasks on the ToDo list!