!!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!! !!!
-------------------------------------------------------------------------------
                                       
                                       
                   B O S T O N   U N I V E R S I T Y
                                       
                      Computer Science Department
                                       
                          C O L L O Q U I U M

                        Controllable Visual Cues: 
            Images as Sensory Signals in Complex Control Systems

                            Stefano Soatto

		         Washington University 


 			 Wednesday, April 28th 
                                       
                                11:00am
                       (Coffee served at 10:45am)

                                       
                         Seminar Room / MCS 135
                                       

-------------------------------------------------------------------------------

    Any imaging system - such as a video-camera, the human eye or a
    telescope - entails a map of the 3-D environment onto the 2-D surface
    of an imaging sensor. Images of such a map are characterized by a
    loss of information along one spatial dimension.

    The goal of a visual system (whether animal or artificial) is to
    exploit ``cues'' in 2-D images to retrieve a 3-D representation of the
    environment. Such a representation is vital in accomplishing spatial
    control tasks such as moving within the environment and interacting
    with it. 

    In order to achieve this goal, 2-D images can be combined with
    a-priori assumptions about the scene. For instance, assumptions about
    the reflectance properties of a scene allows associating variations in
    image brightness (shading) to its the 3-D shape.  Assumptions about
    the local statistics of photometric patterns allows associating
    scene's ``texture'' to its shape. Shading and texture are just two
    examples of pictorial cues, i.e. cues which are associated to one
    single still image. All such cues are intrinsically ambiguous
    (i.e. they generate illusions) in that the a-priori assumptions made
    can never be validated.

    Controllable cues, on the other hand, are not present in one single
    image. Rather, they encode variations among different images of the
    same scene, where the change in the imaging process can - to a certain
    extent - be controlled. For instance, when we change the viewpoint we
    have "Parallax Cues" (Stereo and Motion), and when we change the
    geometry of the imaging device we have "Accommodation Cues". Unlike
    pictorial cues, controllable ones can - under suitable conditions -
    give unequivocal information about the 3-D structure of the
    environment.

    The analysis of controllable cues involves a relatively unexplored
    blend of Differential Geometry, Dynamical Systems and Stochastic
    Processes.

    We will present algorithms for estimating models of the environment
    using controllable cues with applications to 3-D modeling, image-based
    rendering, Human-Computer Interaction, autonomous guidance and
    endoscopic surgery.

    Host: Stan Sclaroff (sclaroff@cs.bu.edu)

-------------------------------------------------------------------------------
For colloquium info, including directions, see http://cs-www.bu.edu/colloquium
-------------------------------------------------------------------------------