Structural correspondence as a contour grouping problem

We present a novel viewpoint which approaches the structural correspondence across an image stack in the 3D space as solving a contour grouping problem. Finding 3D cellular tubes becomes finding closed contours. We derive grouping cues between cells in adjacent slices based on their ability to relate in the 3D space. Those that form a long 3D tube in the space become the most salient contour, while those of shorter lengths become less salient. In the spectral graph-theoretical framework for contour grouping, such a separation by the contour length is reflected in complex eigenvectors of different magnitudes, from which these 3D tubes of varying lengths can thus be extracted, obviating the need for identifying missing correspondences.