Motion compensated intermediate viewpoint image interpolation.

  This is a group project by Mai Vu and Michal Smulski.

As a part of the Light Field Camera project at Stanford, arrays of
digital sensors are being built. To make smooth motion pictures from
these array cameras, one of the challenges is to interpolate images at
intermediate view points between the cameras from the images captured
by these cameras. Since different camera capture the same scene but from
different angles and each camera parameters are different, this creates
variation in the intensity, the baseline and orientation of the images
captured by these cameras.

We propose to estimate the motion vectors between two images and use
this information to interpolate the intermediate viewpoint image. We
currently investigate the following methods for motion estimation:
 
  1. Phase correlation: this method estimates the relative shift between
  two image blocks by means of a normalized cross-correlation function
  computed in the 2-D spatial Fourier domain. An advantage of this
  method is its relative insensitivity to changes in illumination,
  assuming the changes are caused by shifts in the mean values or
  multiplication by a constant.
 
  2. Apply gamma adjustment to each images so that the intensity of the
  images are within the same range and apply block matching method to
  find motion vectors.
 
  3. Use multiple baseline stereo matching method where matching is
  performed by computing the sum of squared-difference values with
  respect to the inverse distance rather than the disparity. This
  technique has been shown to produce a unique and clear minimum at the
  correct matching position even when the underlying intensity patterns
  of the scene include ambiguities or repetitive patterns, thus reducing
  the false matches [4].
 
We are going to investigate theses three algorithms and then choose one
for actual implementation. The motion estimation system will be built
using Matlab. A data set will be created using the current camera setup,
alternatively we may obtain stereo images from the Computer Vision
web-sites. Final part of the project will involve testing the
algorithm. Two possible tests are considered. Firstly we would inspect the
motion vectors to confirm that they correspond to actual data. The final
test will involve using OpenGL to generate a new image using disparity map
and original images.

  References:
 
  1. A.M.Tekalp, "Digital Video Processing", Prentice Hall PTR 1995.
  2. H.Y.Jang, et.al., "Fast Interpolation Technique on Epipolar Plane
  Image Using Phase Correlation", IEEE International Symposium on Circuits
  and Systems, June, 1997.
  3. D.V. Papadimitriou, T.J.Dennis, "Stereo disparity analysis using
  phase correlation", Electronic Letters, Sep 1994.
  4. M. Okutomi, T. Kanade, "A Multiple-baseline Stereo", IEEE Trans. on
  Pattern Analysis and Machine Intelligence, Vol 15, No 4, Apr, 1993.