Background

In the natural world, there is a phenomenal variety of different eye designs --- high resolution, camera-type eyes consisting of a lens and a retina; lower resolution insect eyes made up of thousands of nearly identical components; and the exotic eyes of the stalk-eyed flies, or the stomatopods. This biological diversity has been mimicked in the computer vision community --- the past decade has brought the introduction of very wide angle cameras, catadioptic and omnidirectional cameras, multi-camera networks, and plenoptic cameras. Because these different camera designs sample the local visual field differently, they are suited to different tasks and environments.

Current Work

I am exploring the extremes of the space of camera designs. Below is one example, the "fiber optic camera". This camera (and nearly all new camera designs) fit within the generalized imaging model proposed recently by Grossberg and Nayar. The drawbacks of this camera design are (1) that it is time consuming (not difficult, just time consuming) to calibrate, and (2) that since it does not approximate a central projection (pinhole camera type model), new algorithms are necessary for the motion estimation process. The ego-motion constraints are defined in a paper presented at Omnivis 2002.

These ego-motion constraint define an error function. This function is minimized to determing the ego-motion parameters. This problem is of the same form for all different camera varieties. Considering the Fisher Information Matrix of this error function makes it possible to give quantitative comparisons of different camera designs in terms of their ability to estimate ego-motion. It also gives a mechanism for optimizing the design of a camera system for a particular environment --- or reverse engineering biological eye designs to determine for what environment they are adapted.

This "Fiber Optic Camera" is one extreme of a highly non-standard mechanism for sampling the visual field. This camera can sample the visual field either as a uniform sampling of a hemisphere,

or, in other sampling patterns, which may be optimal if there is prior knowledge of the possible parameters of the motion, (in this case a robot that can only turn around the vertical axis and move straight forward).

Here is the picture of the bottom of the fiber optic bundle, the camera captures this image, but only need to report the intensity of each fiber optic block, because each fiber strand has (effectively) constant intensity.