Outlier reduction for statistical shape analysis

In order to accelerate the statistical shape modelling and enhance the quality of the model obtained, we propose to detect these "outlier'' shapes, and exclude them in training the shape model. The idea to solve this shape outlier detection problem is firstly to decompose the shape using spherical harmonics basis functions, and thus convert the shapes into series of coefficients. To better describe the distribution of target shapes, we propose a new one-class classifier, which uses a hyperellipsoid to enclose the target shapes. Using this outlier filtering scheme, the outlier shapes can be automatically separated from target shapes and the resulting shape model could thus be improved in terms of compactness and specificity.