Algebraic Polyhedral Constraints and 3D Structure from Motion

In [1] Murray et al. describe a motion processing system, ISOR, which is able to recover the 3D motion and structure of polyhedral objects from an image sequence and goes on, where possible, to recognize the object as one from a database of object models. The system performs a 'bottomup' pass through a vision processing hierarchy in the four stages: (i) Low level Compute visual motion at intensity edgels in a sequence of time-varying imagery; (ii) Segmentation Segment the edgels (and thereby visual motion) into groups lying on the same straight edges in the image; (iii) Structure-from-motion (SFM) Compute the 3D structure and motion of the partial wireframe of which the linked straight edges in the image are the perspective projection; and (iv) Recognition — Match the 3D partial wireframe to a complete wireframe stored in a database of object models.