Document Sample
interpolation Powered By Docstoc
					                      CS 563 Advanced Topics in
                             Computer Graphics
       View Interpolation and Image Warping

                               by Brad Goodwin

Images in this presentation are used WITHOUT permission
                                   Over View

 General Imaged-Based Rendering

 Interpolation

 Plenoptic Function

 Layered Depth Image (LDI)
 Image Based Rendering (IBR)
      Composed of photometric observations
      Mix of fields (photogrammetry, vision, graphics)
      Texture mapping
      Environment mapping
      Realistic surface models
      Uses from virtual reality to video games
      Just render the 3D scene?
      Judge results?
 Different types of rendering using different amounts of geometry

 Morphing
   interpolating texture map and shape
 Generation of a new image is independent of
  scene complexity
 Morph adjacent images to view between
 based on viewpoints being closely spaced
 Uses camera position, orientation and range
  to deteremine pixel by pixel
 Images pre computed and stored as morph
                                  About this method

 Method can be applied to natural images
   Only synthetic were tested with this paper
   Of course this paper was in ’93 so hopefully
    someone’s tested them by now
 Only accurately supports view independent
   Others could be used on maps but they are
                                  Types of Images

 Can be done with natural or sythetic images
 Sythetic
   easy to get the range and camera data

 Natural
   Use ranging camera
   Computed by photogrammetry or artist
                                           General Setup

 Morphing can interpolate different
     Camera position
     Viewing angle
     Direction of view
     Hierarchical object transformation
 Find correspondence of images
   Images arranged in graph structure
                             Find correspondence
 Usually done by animator
 This method
   Form of forward mapping
   uses camera and range to do it
 Cross dissolving pixels(not view-independent)
 Done for each source image
 Quadtree compression
   Move groups of pixels
 Scene moves opposite camera
 Offset vectors for each pixel (“morph map”)
   Small change more accurate when interpolated
 Sampled every   Offset vectors
  20 pixels
                                  Overlaps and holes

 Overlaps
   Local image contraction - several samples move
    to the same pixel in interpolated image
      Perpendicular to oblique
 Holes
   Show when mapping source to destination
   Background color
   Interpolate four corners of the pixel instead of
    center (filling and filtering)
   Interpolate adjacent offset vectors
   Or if part seen in interpolated but not source
                                      Block Compression
 Pixels ten to move together so block compression
  algorithm is used to compress morph map.
 Related to image depth complexity
    High complexity low compression ratio
                              View independent Priority

 Established to determine points that are viewable
 Pixels are ordered from back to front based on Z-
  coordinates established in morph map
 Eliminates need for interpolating the Z-coordinates of
  every pixel and updating the Z-buffer in the interpolation

 Virtual Reality
 Motion blur
    Uses super-sampling of many images
     computationally which is expensive thus
 Reduce cost of computing a shadow map
    Only for point light sources
 Create 3D primitives without creating 3D
                                     Plenoptic Modeling

 The Plenoptic function
   Latin root       plenus – complete or full
                     optic - pertaining to vision
   Parameterized function for describing everything
    that is visible from a given point in space
   Used as a taxonomy to evaluate low-level vision
   Adelson and Bergen postulate
     “…all the basic visual measurements can be considered to
       characterize local change along one or tow dimensions
       of a single function that describes the sructure of the
       information in the light impinging on an observer.”

 azimuth and elevation angle

 Set of all possible environment maps for a
  given scene

 Specify point and range for some constant t

 A complete sample can be defined as a full
  spherical map
                             Plenoptic Modeling

 Claimed that all image-based rendering
  approaches are just attempts to create a
  plenoptic function with just a sampling of it
 Set up is the same as most approaches
 Set of reference images which are warped to
  create instances of the scene from arbitrary
  view points
                           Sample Representation
 Unit sphere
   Hard to store on a computer
   Example of all distorted maps
 Six planar projections of a cube
   Easy to store
   90 degree face requires expensive lens system to
    avoid distortion
   Oversampling in corners
 Have to choose Cylindrical
   Easily unrolled
   Finite height :problems with boundary conditions
   No end caps
                                  Aquiring Cylindrical
 Get the projections is simple
 Tripod that can continuously pan
 Ideally camera’s panning motion should be
  exact center of tripod
   When panning objects are far away slight
    misalignment is tolerated
 Panning takes place entirely on the x-z plane
 Both images should have points within each
 Find the projection of the output camera on
  input cameras image plane
 That is the intersection of the line joining the
  two camera locations with the input camera’s
  image plane
 Line joining the two cameras is the epipolar
 Intersection with the image plane is the
  epipolar point
 Map image point to output cylinder
    Same techique for comparing points used with face mapping
     from last week
                                 Layered Depth Images
 Paper presents some methods to render multiple
  frames per second on a PC
 Sprites – are texture maps or images with alphas
  (transparent pixels) rendered onto planar surfaces
 One method warps Sprits with Depth
    Warps depth values and uses this information to add
     parallax correction to a standard sprite renderer
    Single input camera
    Contains multiple pixels along each line of sight
    Size of representation grows linearly with the depth
     complexity of the scene
    Uses McMillan’s warp odering algorithm because data is
     represented in a single image coordinate system.
   Chen S E and Williams L, "View Interpolation for Image
    Synthesis", Proc. ACM SIGGRAPH '93 McMillan L, and Bishop,
    "Plenoptic Modeling: An Image-based Rendering System", Proc.
   Shade, Gortler, He and Szeliski, "Layered-Depth Images", Proc.
   McMillan L. and Gortler S,"Applications of Computer Vision to
    Computer Graphics: Image-Based Rendering - A New Interface
    Between Computer Vision and Computer Graphics, ACM
    SIGGRAPH Computer Graphics Newsletter, vol 33, No. 4,
    November 1999
   Shum, Heung-Yeung and Kang, Sing Bing, A Review of Image-
    based Rendering Techniques, Microsoft Research
   Watt, 3D Graphics 2000, Image-based rendering and phto-
    modeling (Ch 16)