On Properties of Active Shape Models

Contrary to many other deformable models Active Shape Models (ASM) represents a general way of performing non-rigid object segmentation. Shape variation is extracted from a training set by applying principal component analysis to point distribution models, rather than hand crafting a priori knowledge into the model. In this paper we investigate di erent properties of ASM. Topics treated are the generation of plausible shapes, tangent space transformation and model to image tting assisted by statistical models of gray level variation in the training set. Finally a method for automatic initialization and a comparison of four model to image tting methods are presented. The initialization part indicates that completely automatic segmentation could be done by ASMs. The comparison part shows an improved t for model to image t methods based on gray level variation in the training set.