<grumble>

My understanding was that it could back-calculate the retinal map.

Alas, I was indeed wrong.

For those who don't mess with this OpenCV, what this does is bring the colorspace and attributes of an image to something that the eye would see.

camera picture + retinal filter = simulation of how the eye would see the picture captured by camera.

Hah, I was about to get mad at myself for not realizing retina biometrics were built in to OpenCV.

As someone who has worked a bit using OpenCV to detect eye features, I can say pretty confidently that it isn't THAT easy. OpenCV can help you start off by easily being able to detect that there is an eye in frame, but I think you are on your own beyond that.

My general approach was to use haarcascades to detect that an eye was in frame. Then you have to isolate the iris, and pupil. You can use Hough Circles or the Daugman algorithm for segmentation. I got a reasonable result using the Daugman algorithm, but Hough circles seemed to be more erratic. I had a huge problem with reflections though, which I think it is known that the Daugman algorithm suffers from. You pretty much have to take the picture of the eye in a controlled setting where you flash a light in the subjects eye such that the reflection in the eye is a small isolated circle. Even then I wasn't always to get a correct result. I think I may have not implemented the algorithm perfectly though.

I never got farther than that, but even if you are able to capture the eye perfectly you then have to actually build up a model of the persons features and then be able to compare it.

I would be interested if anyone else had better luck than I did.