In [106] Spoerri and Ullman claim to segment the images before flow is found, using various flow discontinuity tests. (It seems unlikely that this is achieved, as edge-based motion is used to drive the local discontinuity tests.) The sort of tests which are used are bimodality in local normal flow histograms and appearance/disappearance of thin image segments, requiring fine image texture. Only very simple images are used for testing. In [81] Overington utilizes the normal components of flow computed at edges (using gradient-based techniques) to find discontinuities in the normal flow component. The discontinuities are used to detect moving objects in a scene, taken from a static camera.
In [120] Waxman and Duncan discuss the use of stereo and motion for recovering information about the world. In their experiments they tracked dots (painted onto objects) by assuming that image motion was less than dot separation. They attempt to segment spatially discontinuous scenes by finding boundaries of ``regions of analyticity'' in the flow. Even with the artificially aided (and therefore hopefully accurate) way of recovering optic flow used, the segmentation boundaries are poor.