Data compression of stereopairs

If no data compression scheme is employed, twice as many bits are required to represent a stereopair as are required to represent a single image; this can make using stereopairs impractical. The goal of this research was to develop ways to exploit the stereo correlation to reduce the number of bits required to represent a stereopair. This dissertation addresses the stereopair compression problem both theoretically and practically: first, the problem is treated within the context of rate-distortion theory; second, several techniques for compressing stereopairs are discussed; and finally, a method for assessing the quality of encoded stereopairs is presented.

A discrete-time, discrete-alphabet, stationary, ergodic random process model of a sequence of stereopairs is developed. This model is used to prove that the coder/decoder structure "code right picture, code left given encoded right; decode right, decode left given decoded right" is inherently sub-optimal with respect to achieving the rate-distortion limit.

Several techniques designed to minimize the mean square error between a stereopair and its reconstruction are presented. The most successful technique is based on disparity-compensated transform-domain prediction of the left picture from the encoded right picture. Using this technique, stereopairs can be coded with up to 20% fewer bits than are required when the left and right pictures are coded/decoded independently. A psychophysically based technique, called mixed-resolution coding, is also introduced. When a lowpass-filtered picture is shown to one eye, and a high-quality picture is shown to the other eye, the resulting stereo percept closely resembles that of the high-quality right-eye picture. The lowpass-filtered picture can be subsampled to achieve compression. The number of bits required to represent the left picture can be reduced by a factor of 16 using this method.

In many applications compressed stereopairs are viewed to perform manipulation tasks. In these cases the crucial question is how an observer's ability to perceive depth in a compressed stereopair compares with his ability to perceive depth in an uncompressed stereopair. A method for making this comparison is presented. We found that observers generally perceived objects to be further away in compressed stereopairs than they did in uncompressed stereopairs.