Design of orientation assessment functions for gestalt-grouping utilizing labeled sample-data
Psychological evidence is given that perceptual grouping is an important help for various visual tasks. Object recognition and land use classification from remotely sensed imagery is an example. In machine vision, such a grouping process can be implemented by coding Gestalt laws such as proximity, symmetry, or good continuation. Since geometric relations are rarely fulfilled exactly, soft membership functions are utilized called Gestalt assessments. Hierarchical grouping is possible on increasing scales. Such an approach to hierarchical Gestalt grouping is modified in this paper. In its original form, the approach uses rather heuristic default assessment functions, which are a possible choice as long as no labeled example data are given. The assessment functions can be parameterized so as to improve the perceptual grouping, guiding it by the Gestalten salient to human perception. To this end, we use orientation statistics from the publicly available data set given for the ICCV symmetry recognition competition 2017. Also, with a particular recognition task at hand, labeled example data can serve as the desired foreground. Here we use the ground-truth layer for buildings of the Vaihingen benchmark of the ISPRS. A mixture distribution containing two von Mises-distributions and the uniform component for the clutter in the background is fitted using expectation maximization.