Simon Baker, Iain Matthews (2004)
- Usual approach to image alignment is gradient descent
- Different ways: additive (additive increment to parameters) or compositional (estimate incremental warp that is then composed with current warp estimate)
- Another difference: Gauss-Newton, Newton, steepest-descent or Levenberg-Marquardt approximation in each gradient descent step
- All are orthogonal choices (any combination possible)!
- Choice depends on:
- Most noise in template (a) or input (b) image
- Need for an efficient algorithm
- Best choices:
- (1a): forwards algorithm
- (1b): inverse algorithm
- (2): inverse-compositional Gauss-Newton or inverse-compositional Levenberg-Marquardt (probably also best options for noisy templates!)