Several problems in computer vision and graphics such as pattern
classification, noise removal, and texture synthesis have been
approached using function approximation. In practice, however,
there is usually not enough training data to approximate the
function continuously, and the data that is available is sparse.
The function to be learned is represented as piecewise constant
and point neighborhoods present in the training data are used to
approximate the function in areas where no data is available. We
present a general algorithm and show how to apply it to several
challenging computer graphics and vision problems. We use the
algorithm to learn simple video processing operations, to make
videos appear to be moving versions of paintings, and to make
views of 3D models appear to be photorealistic.
Please note: sequences have been compressed to reduce download time.
As such, some contain compression artifacts.
3D surface material property transfer