Perceptual grouping is a well studied area in visual psychophysics and offers a principled, general way to study vision in sighted animals and humans as well as machines. Computational approaches in the past however have often been hampered by complexity issues and brittleness in the presence of clutter and noise.
Especially the reliance on tuning parameters renders many approaches impractical for real world applications.
This work aims to address complexity and robustness issues by proposing an incremental processing scheme for the perceptual grouping of edges, where the only parameter is runtime. This allows interrupting processing at any time, returning the most significant perceptual groups that could be found up to that point and leads to graceful degradation with increasing amounts of noise or clutter.
We furthermore propose a probabilistic measure of visual significance based on the principle of non-accidentalness. This significance measure is used to guide grouping of convex contours as well as a relative depth ordering of contours.
Varying the significance measure of edges, for example based on regions of interest, allows to focus attention on specific parts of the scene, which will subsequently be allocated a larger share of the available processing time.
Experiments were carried out on a wide range of real world images with varying scene content and complexity. For detection of convex contours we could show that identifying candidate edges and junctions for grouping can be performed in runtime linear to the number of edges. More significant contours typically popped out faster with less significant ones appearing as runtime progresses. We demonstrated how two attentional mechanisms, based on regions of interest and colour, lead to faster detection of objects of interest. Relative depth ordering of contours based on energy minimisation in a Markov Random Field was presented as an initial form of finding a globally consistent scene interpretation and was shown to work for scenes of limited complexity.