Determining 3D information together with correspondence from a sequence of orthographically projected optical flows-integrated approach to structure-from-motion issue

Structure-from-motion is considered to consist of two consecutive stages: the correspondence stage followed by the 3D recovery stage. So far the two stages have been mostly studied separately, although they are closely related to each other. We integrate the two stages. We show the benefits of the integrated approach: the correspondence problem can be solved more reliably in this approach than in the separate approach; the recovery problem can be solved more efficiently; the inaccuracies in the given optical flow, 2D velocities, can be improved by feeding back the results of recovery. We focus on perceptual phenomena `biological motion´ studied by (Johansson, 1975). A computational theory has been presented by (Kato et al., 1992). The theory, however, is concerned only with the 3D recovery stage. We introduce the overall scheme of the integrated approach consisting of the correspondence stage followed by the 3D recovery stage. In the correspondence stage, the candidates of point-to-point correspondence between consecutive frames are proposed according to a priority based on a certain Gestaltian measure. When in the 3D recovery stage, each candidate is examined according to priority if it satisfies the constraints of a moving articulated object. The result is fed back to the correspondence stage to improve the accuracies of optical flows. Then the improved optical flows lead to better recovery, and so on. The results of simulation studies are shown to demonstrate the validity of the integrated approach