Increasing the contrast of just one eye's image degrades stereothresholds; this phenomenon is referred to as the stereo contrast paradox. In experiment one, this paradox was found to be absent in dynamic random-element stereograms; thresholds were simply limited by the lower of the two eyes' contrasts. In experiment two, in which narrowband Gabor targets were used, the paradox was found to be strongest at relatively low spatial frequencies (1 cycle deg-1). As spatial frequency was increased, the paradox gradually disappeared. At relatively high spatial frequencies (5 cycles deg-1), thresholds were generally limited by the lower of the two eyes' contrasts, as was found for the dynamic noise targets. These results demonstrate the interactions of spatial frequency and contrast in binocular image combination and yield clues as to the different roles which high and low spatial frequencies may play in stereopsis.

In three experiments, asymmetries between the processing of crossed and uncrossed disparities were investigated. The target was a luminance-defined circle concentric to a fixation mark, viewed stereoscopically on a computer monitor for 105msec, Fifteen disparities were presented according to the method of constant stimuli. Observers indicated the apparent direction of target depth relative to fixation. All experiments measured both the accuracy and latency of this response. Experiment 1 showed fewer errors and shorter reaction times for identifying crossed disparities. Experiments 2 and 3 replicated Experiment 1 and also showed that observers may often perceive a target in the direction opposite that prescribed by the disparity information. Wepropose that the asymmetries and reversals result from differences in computation of sign, not of magnitude. This notion is consistent with a scheme of continuous disparity tuning and accounts for such asymmetries and errors without positing disparity pooling mechanisms.A fundamental issue in the study ofstereopsis concerns the manner in which retinal disparities are processed by the visual system. The two most popular schemes involve either a small number of relatively independent channels (or pools ofdisparity detectors), or a large number ofdistinct yet continuously distributed disparity detectors.The first scheme was proposed by Richards (1971) on the basis of experiments that revealed anomalous individual differences, or "stereoanornalies," in disparity processing. Richards found that over half of the observers he tested were insensitive to either crossed or uncrossed disparities in local stereograms. In addition, he observed that, for any observer, two very different disparities of the same sign could signal the same depth percept. Some observers even exhibited an apparent reversal of the disparity sign information: The entire range ofdisparities (crossed and uncrossed) elicited depth in only one direction. Based on these findings, Richards posited at least two distinct and independent pools of integrated disparity detectors (crossed and uncrossed). That is, Richards presumed that, since some observers appeared insensitive to the entire range of disparities of a particular direction, the disparity detectors for that direction were pooled into a single (malfunctioning) mechanism. Moreover, the pheThis work was supported by NIH/NEI EY 10303 and AFOSR F49620-93-I-0307. We wish to thank S. B. Stevenson for his helpful comments regarding the results and R. P.O'Shea, R. Patterson, and two anonymous referees for their helpful comments regarding an earlier

Constraints on binocular matching were investigated by comparing the thresholds for interocular correlation in random element displays for human and model observers, with element density manipulated as a parameter. The models consisted of ideal decision rules operating on the entire stimulus, only on the edges in the stimulus, or only on the sparse minority elements in the stimulus. The results indicate that the human visual system selectively attends to the stimulus edges or to the sparse elements under most circumstances. Efficiencies (human or model) were highest at very low element densities (approximately 20%) and decreased with increasing element density with a log-log slope of -0.5, indicating that dynamic random element stereograms at the traditional 50% element density are vastly undersampled.