Fine point alignment

Locally shift points using an elastic surface metaphor. Identify plausible correspondences between cameras and deform the points so as to maintain the regular 2D pixel layout but to allow depth modifications. This is similar to ICP for adjusting camera pose, except it also allows local modification of point locations by the adjustment of depth values in addition to pose adjustment.

Potential energy is used, the points are held at a potential energy equilibrium in a pixel lattice, but are attracted by some force towards an identified correspondence. The algorithm has two steps: calculate energies, and perform the movement. The entire process is then iterated.

A further point: normals that are close should match, but those that are also close to the cameras "normal" should be weighted substantially more strongly? We are using regular 3D weighting function, but should this not be adapted to an "egg" shape to match the confidence or covariance of the point? Depends if our registration is more or less accurate than our stereo depth estimates.