Robust outliers detection in image point matching

Classic tie-point detection algorithms such as the Scale Invariant Feature Transform (SIFT) show their limitations when the images contain drastic changes or repetitive patterns. This is especially evident when considering multi-temporal series of images for change detection. In order to overcome this limitation we propose a new algorithm, the Affine Parameters Estimation by Random Sampling (APERS), which detects the outliers in a given set of matched points. This is accomplished by estimating the global affine transform defined by the largest subset of points and by detecting the points which are not coherent (outliers) with the transform. Comparisons with state-of-the-art methods such as GroupSAC or ORSA demonstrate the higher performance of the proposed method. In particular, when the proportion of outliers varies between 60% and 90% APERS is able to reject all the outliers while the others fail. Examples with real images and a shaded Digital Elevation Model are provided.