When a line sensor was regularly curved, the mean plane corresponding to
the line of sights of all pixel was slanted too much and did not contain
the origin anymore.

Near tile boundaries, the algorithm sometime failed to find intersection
points.

This is done by dumping its state into a file, thus avoiding costly
initialization.

I now ned about 56s for 2000x2000 inverse localization for 3 sensors on
my laptop (Intel Core i5-3320M, 2.60GHz, quad-core, linux, 8GB RAM,
mono-thread test). This means more than 210000 pixels/second.

It's fast!

When ground points are close to first or last pixel, we need to be able
to compute angles sightly after line end, so a small 5 pixels margin has
been added in the search algorithm.