Cite as:
Ohshiro M, Shioiri S, Yaguchi H, 1998, "A model of visual segmentation for images with T-junctions" Perception 27 ECVP Abstract Supplement

A model of visual segmentation for images with T-junctions

M Ohshiro, S Shioiri, H Yaguchi

Occlusion cues such as T-junctions in an image are useful to perceive a 3-D world from 2-D retinal images. However, typical algorithmic methods for computer vision have difficulty in analysing the images with occlusion in the way human vision does. We propose a model for detecting the depth relationship between overlapping surfaces based on T-junctions. Our model contains four stages. In the first stage, edges of different orientations are detected separately to make an orientation map by spatial filtering and calculation of local energies with filters similar to cortical simple and complex cells (Gabor filters with twelve different orientations). In the second stage, T-junctions are detected from the orientation map by finding the regions where two perpendicular (or close to perpendicular) edges form the shape of a T. In the third stage, the two orientation components at each T-junction are separated in different depth planes on the basis of the occlusion relationships between the two (the leg of a T is assumed to be behind the crossbar of the T). In the fourth stage, the occluded edges are interpolated by interpolation filters. The interpolation filters are also a set of Gabor filters with a larger space constant. They operate on each of the orientation maps with a weight that depends on the difference between the preferred orientation of the filter and the orientation of the map. This operation generates responses at a point where there is no edge information but there are edges directed to the point in the neighbourhood. The model has been successfully applied to monochromatic images.

These web-based abstracts are provided for ease of seaching and access, but certain aspects (such as as mathematics) may not appear in their optimum form. For the final published version of this abstract, please see
ECVP 1998 Abstract Supplement (complete) size: 2107 Kb