Alex Wade sent me the following bit about a competition he is organizing. There is a $1000 prize plus an invitation to present the method at the OSA Fall Vision Meeting, 2011.
I'd like to draw your (and your readers') attention to a competition that we are running as part of the Optical Society of America:
it might be of interest to you. Illuminant recovery has a 'sparse' kind of feel because it is generally assumed that although spectra appear complex, they are actually well modeled by combinations of 3 well-characterized basis functions.
I looked a little bit into the problem and ask the following to Alex:
... I have a probably very dumb question with an obvious answer: What is the linear transform between the spectral component of a hyperspectral images (several spectral bands) into the L, M, and S coordinates at each image pixel, i.e. what is the weighting used to compute the LMS coordinates ?Cheers,
To what Alex kindly responded with
....The spectral absorption characteristics for the human cones are well known and published. You can therefore compute individual cone absorptions by convolving the cone spds with the illuminant spectrum. This process is fully described in the sample programs (available in both Matlab and Python) which also demonstrate how to read and write the data format we prefer.The goal of the contest:
We have provided a set of 10 test images. The contest is to estimate the relative illuminant spectrum for each of these images. The winning entry will be the one that has the minimum estimation error, averaged across the 10 images.
In the rules, I noted the following two items of specific interest:
* You can use any method you like, other than hacking into our computers to find the actual illuminant data. That is, computer algorithms, human assisted algorithms, guesses based on intuition, crowd sourced solutions, etc. are all fine. We will not score entries so rapidly as to make parameter search on the returned score a particularly feasible approach, although information gleaned from the score obtained via submissions may be used.
* The winner will be the top scored entry as of midnight, July 14, 2011. In the case of a tie, the cash prize will be split evenly and we will do our best to make fair arrangements for the invited Fall Vision Meeting presentation.
My emphasis, at the very least, they are aware of overfitting issues such as the one that occur in the Matlab contest. I also noted in the image generation section, the following tidbit:
We note here only that the images were rendered using perspective projection and that the parameters were set to simulate the effect of mutual illumination bet
if mutual illumination is on, there might be some nonlinearity here. But eventually, it looks some sort of spectral unmixing and blind deconvolution. Enjoy!