Several approaches have been proposed for this minimization through years of
studying. The first idea was use an additive approach, in which ∆pi
and ∆λi
were
calculated as linear functions of the error image and then shape parameter p and
appearance λ were updated as pi
← pi
+ ∆pi
and λi
← λi
+ ∆λi
, in the i-th iteration.
Although convergence can occur sometimes, the delta doesn't always depend on
current parameters, and this might lead to divergence. Another approach—which
was studied based on the gradient descent algorithms—was very slow, so another
way of finding convergence was sought. Instead of updating the parameters, the
whole warp could be updated. This way, a compositional approach was proposed
by Ian Mathews and Simon Baker in a famous paper called Active Appearance Models
Revisited. More details can be found in the paper, but the important contribution it
gave to fitting was that it brought the most intensive computation to a pre-compute
step, as seen in the following screenshot: