Animation using Optic Flow - Ian Webb's Research

My research is in the area of computer graphics, in particular animation using optic flow. In April 1998 I completed an MSc degree in the Computer Science Department at the University of Cape Town. The degree was awarded with distinction.

butterfly Project title and summary

Supervisor:
Edwin Blake

Project Title:
`An extension of optic flow analysis for the generation of computer animated images.'

Or, in English, using some mathematics to find faster ways of doing computer animation, using information already present in the current frame to produce the next frame, and therefore avoiding unnecessary work. A philosophy worth looking into.

The intention is to make use of optic flow analysis to extract the transformations between frames in a synthetic animation. This allows us to do image-based rendering, reusing information from previous frames rather than a brute-force approach where each frame is rendered independently.

This approach makes use of available information such as the relative motion of objects and viewer (which occur in 3D space) and the projection into 2D screen space. From this information a transformation in 2D can be derived which describes how the image changes from one frame to the next. This description may be incomplete due to occlusions, but in general it is valid over large continuous areas of the image.

Relating animation to computer vision and image sequence compression

Although our work falls into the category of image-based rendering, related work has mostly been done in the fields of computer vision (using optic flow to extract 3D structure from a video sequence) and image sequence compression (MPEG-like compression schemes which use optic flow information to encode the transformations between frames).

Animation is in some ways an inverse of computer vision - instead of extracting 3D struture and motion from the video sequence, we want to generate that sequence knowing the structure and motion. The mathematics of optic flow analysis is identical in both cases.

Since we are dealing with a synthetic animation rather than a video sequence, we have full knowledge of the optic flow field, without issues such as noise or contour extraction which make establishing pixel correspondences a hard problem in computer vision.

In image sequence compression, information from frame N and frame N+1 is used to extract and encode the transformation between frames. We are interested in generating the transformation from frame N to frame N+1 and using it to produce frame N+1 using image transformations.

Theoretical background

Experiment

References

Ian Webb