PDF version of this document

Model Evaluation

Shape models were previously evaluated using the two measures named specificity and generalisation ability (generalisability in short). Model complexity measures were initially investigated by Kotcheff [18] and further use was made of them in the work of Davies where shape models were optimised.

The idea behind this reverts back to fundamentals where models of visual forms in fact describe clouds in a high-dimensional space (visualised in Fig. 2). A model is essentially a descriptor of a volume in space - a mean point with knowledge about its extent of variation in the different directions.

If each model is in fact a simple cloud in space, models can be compared by measuring the overlap of clouds analytically. Models are evaluated by comparing them with respect to the training set, which are interchangeably just a set of points in that space. The models are best represented by a large collection of syntheses which are derived from them.

Fig. 2. Training set and model synthesis in hyperspace