Shape restoration for robust tangent principal component analysis

Shape outliers can seriously affect the statistical analysis of the shape variations usually performed by the Principal Component Analysis PCA. This paper presents an algorithm for outliers detection and shape restoration as a new strategy for robust statistical shape analysis. The proposed framework is founded on an elastic metric in the shape space to cope with the nonlinear shape variability. The main contribution of this work is then a formulation of a robust PCA which describes main variations associated to correct shapes without outlier effects. The efficiency of this approach is demonstrated by an evaluation carried out on HAND-Kimia and HEART-Kimia databases.