We show a simple and efficient way for rendering arbitrary views from
so-called free-form light fields, employing a convex free
form camera surface and a set of arbitrarily oriented camera
planes. This way directionally varying real-world imagery can be
displayed without intermediate resampling steps, and yet rendering of
free form light fields can be performed as efficiently as for
two-plane-parameterized light fields using texture mapping graphics
hardware. Comparable to sphere-based parameterizations, a single free
form light field can represent all possible views of the scene without
the need for multiple slabs, and it allows for relatively uniform
sampling. Furthermore, we extend the rendering algorithm to account
for occlusion in certain input views. We apply our method to synthetic and real-world datasets with and
without additional geometric information and compare the resulting
rendering performance and quality to two-plane-parameterized light
field rendering.
