T-junctions: Cues for segmentation rather than for occlusion

In everyday situations, we experience a world of whole objects, even though many of these objects are partly occluded by each other. The research field of visual pattern completion focuses on the question of how a partly occluded object is completed amodally (i.e., beyond the visual input) into a whole object. In this research field, a distinction is made between local and global approaches.

Local approaches focus on the junctions where the visible edges of different objects meet. The figure below highlights two so-called T-junctions between, as far as visible, a red rectangle and a blue cross. According to local approaches, T-junctions are cues for occlusion. In the figure below, this implies that the rectangle is predicted to be the occluder, and that the visible edges of the cross are smoothly continued behind the occluder until they meet, which here yields a non-cross as whole object.

Global approaches focus on the shapes of all candidate whole objects, starting from the visible contours (as highlighted in the figure below). For each candidate occluder and each candidate occluded, global approaches quantify the goodness-of-shape and then select the best-of-shape combination. In the figure below, the (already visible) cross has a higher goodness-of-shape than the (partly visible) non-cross, which implies that no occlusion is predicted here.




Does this mean that, according to global approaches, T-junctions are not relevant to amodal completion? The amodal completion model within the global approach of SIT model leads to a balanced answer along the following line of reasoning.

To explain amodal completion, this model considers all candidate occluder+occluded combinations for a given stimulus. For each combination, it first quantifies the goodness-of-shape of the occluder and the occluded separately, namely, by the structural complexity of the simplest SIT code for each shape. The sum of these complexities yields the total shape complexity I_shape. Then, it quantifies the structural complexity of the relative position of the occluder and occluded, yielding the positional complexity I_position. Finally, the combination with the smallest sum I_shape + I_position is predicted to be the preferred stimulus interpretaton.

SIT's amodal completion model has shown to have considerable predictive power. Without going into quantification details, the next figure shows its predictions for some stimuli. These stimuli consist of edges that, at their point of contact, form one of four different junction types (as highlighted by the circles). For each stimulus, the "two objects" hypothesis and the "one object" hypothesis are considered, and the boxed complexities indicate the predicted interpretations.




Thus, for instance, the third stimulus is predicted to be seen as one L-shape rather than as two objects with an L-junction at their point of contact. The second stimulus, however, is predicted to be seen as two objects with a T-junction at their point of contact rather than as one T-shape. This implies, inversely, that a T-junction is a cue for segmentation, that is, a cue for the presence of two objects.

Hence, according to SIT's global approach, T-junctions are relevant to amodal completion, albeit it not as direct cues for occlusion but merely as cues for segmentation. By consequence, whether or not the resulting segments are seen as belonging to partly-overlapping objects, must co-depend on the shapes of the hypothesized whole objects.


For further demos on these issues, see Object versus viewer and Occam versus Bayes

For a full account of SIT's amodal completion model, see Perception 1994
For a multidisciplinary and historical embedding of this occlusion model, see Psychological Bulletin 2000