11/03/2026 2:11 PM - PACES
11/03/2026 2:11 PM - PACES
Aidan Dempster - adempst - 6125 9596
Problem
Find the best match for each pixel in one image to another image. Although they solve it for a single registration where you match one image to another.
Approach
They use an iterative approach where you repeatedly find linear approximations of the error gradients and update to minimize them. This is similar to hill climbing approaches previously used, but more efficient. They weight the gradient by an estimate of the curvature in order to avoid overshooting.
Contribution
The optimization process they use is Newton Raphson which is a novel way of solving image registration. The usage of weighting given the curvature is also a novel contribution. They also use a course-to-fine strategy of iteratively optimizing the position so that it first matches the positions of large blobs and then refines that to higher and higher frequency components. This allows for much greater receptive field.
Evaluation
They guarantee convergence theoretically using their derivations and show that it works on simulated data with a sin wave. They also show convergence on a real test image.
Substantiation
They have pretty weak substantiation, but no worse than other papers we have been looking at. I am convinced that it does converge for at least a test image.