Primitive extraction

There is an important algorithmic step between the input in the form of images in pixel raster formats and, possibly hierarchical, perceptual grouping: the extraction of primitive Gestalten to start with. Often a failure in recognizing a symmetry, which is present in an image either as evident to human observers, or as given by some ground-truth, cannot be blamed on the Gestalt operations. Instead, too much information is lost already in the extraction of the primitives. This chapter lists a set of possibilities, including very simple and fast methods such as threshold segmentation as well as sophisticated automatic feature extraction methods such as scale-invariant feature transform (SIFT), or the maximally stable extremal regions (MSER). It is of course also possible to use machine learning methods for primitive extraction, and the chapter includes some discussion on this topic as well. In particular self-organizing maps are proposed for color images and hyper-spectral images.