Unformatted text preview: owards this end, it is rst shown that two basis views
are su cient to predict the appearance of the scene
within a speci c range of new viewpoints. Second, it
is demonstrated that interpolation of the recti ed basis
images produces exactly this range of views. Finally,
it is shown that generating this range of views is a theoretically well-posed problem, requiring neither knowledge of camera positions nor 3D scene reconstruction.
A scanline algorithm for view interpolation is presented
that requires only four user-provided feature correspondences to produce valid orthographic views. The quality
of the resulting images is demonstrated with interpolations of real imagery. 1 Introduction Despite signi cant advances in 3D computer graphics, the realism of rendered images is limited by handcoded graphical models. Existing techniques for creating 3D models are time intensive and put high demands on the artistry of the modeler. In light of these
limitations, there has been growing interest in the use
of 2D image warping techniques for image synthesis
and animation. The advantage of working in 2D is
that photographs of real scenes can be used as a basis
to create very realistic e ects. A good example of such
an e ect is morphing which combines an interpolating
warp and a cross dissolve of two images to produce
in-between images. Morphing techniques have been
The support of the National Science Foundation under Grant
Nos. IRI-9220782 and CDA-9222948 is gratefully acknowledged. 1 and Y may be any two linearly independent vectors.
Finally, an image is the projection of the visible scene
into the view. An image can be represented as an array of pixels I or a matrix of feature positions I. If S is
a matrix whose columns are the visible homogeneous
scene points then terpolation is explored in Section 5 and a scanline algorithm for view interpolation using minimal correspondence information is introduced in Section 6. Section
7 presents results on real images. 2 Related Work Ullman and Basri 5 demonstrated that new views
can be expressed as linear combinations of other views
of the same scene. Although the focus of their work
was recognition, it has clear rami cations for view synthesis, providing a simple mechanism for predicting the
positions of features in new views. However, their work
does not take into account visibility issues that are crucial to understanding which views can be synthesized.
Chen and Williams 4 described an approach for
view synthesis based on linear interpolation of corresponding image points using range data to obtain correspondences. They investigated special situations in
which interpolation produces valid perspective views,
but concluded that the interpolated images do not in
general correspond to exact perspective views.
Two groups 2, 6 have recently developed image
warping techniques for perspective-correct view synthesis. Under the assumption that a complete pixelwise correspondence is available, it is possible to predict a broad range of views. Several...
View Full Document
- Spring '08
- Computer Graphics, Orthographic Projection, Monotonicity, view interpolation