PDF version of this document

... convergence¹

In the case of active shape models, on the contrary, inspection of nearby structures guided the model so that it better matched the target. Since the model landmarks usually reflect on the location of strong edges, similarity could be well-approximated by the distance between model points and strongest edges in their vicinity. Search along normals to the lines joining these landmarks led to appropriate edges location, provided that initialisation placed the model close enough to the target.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... images²

This reverse process is used, for example, in application where reconstruction of faces or generation of flexible faces is of some value. Real-time animation is another possible extension of this.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... analysis³

Procrustes analysis has proved to be a popular method of shape analysis. The generalized Procrustes procedure was developed by Gower (1975) and has been adapted for shape analysis by Goodall (1991).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... model⁴

The model is constructed using only $n-1$ examples where $n$ is the total number of examples and is then evaluated in accordance with the single left-out example. This procedure is usually performed $n$ times, with a different example put aside at each iteration and averaging is then used to estimate the relevancy of the model.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... folding⁵

For closed curves, convex in particular, such problems are minute yet not negligible. A monotonically increasing reparameterisation function will ensure that points along the curve will at no stage overlap or conflict with one another in some way. It also has the advantage of allowing any number of landmarks to be considered, so resolution constrained could be specified beforehand.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... time⁶

There has been the temptation of optimising several examples from the set in a distributed manner.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... slow⁷

The algorithms currently used for demonstration purposes take 3 days to run, but substantial speed-up is expected soon.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... group-wise⁸

Another advancement in current research is the inclusion of the whole set of data rather than just a one-to-one (pairwise) correlation between data and the current AAM. An investigation of just a couple at a time leads to poorer results in later practical use. That are due to the limited scope of the approach.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... unprecedented⁹

Comment: I am making some risky guesses here (as I often did beforehand). Apologies for any disturbing assumptions that I make to preserve good flow.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.