Objective evaluation of human and computational stereoscopic visual systems

Several computational models have been proposed to account for the abilities of human observers in solving the stereo correspondence problem. Such models are often evaluated by counting the number of correct matches that the model makes when presented with a test stimulus, but this measure is difficult to compare with conventional psychophysical measures of human stereo performance. This paper develops an absolute comparison of human and model stereo correspondence by using the concept of statistical efficiency, through which the performance of both humans and models is compared against that of an ideal observer. Some important assumptions need to be examined before the efficiencies of human and model can be compared because these assumptions affect the statistical performance of models. One assumption involves the use of an epipolar constraint to help constrain the matching problem but it is unclear what the resolution (or "height") of an epipolar line should be. Experimental estimates were obtained of the vertical region over which dot position may be varied without disruption of human stereoacuity. These estimates constrained the effective "height" of the zone above and below an epipolar line. Two models of stereopsis were then compared against human performance: the first (due to Pollard, Mayhew & Frisby, Perception, 14, 449-470, 1985) uses local constraints to solve the correspondence problem; the second is a simple area-based correlation model. For nearly all the stimuli tested, both models show strong similarities with human observers. For one particular stimulus, the efficiency of the correlation model is different from that of humans and of the PMF algorithm. This implies that a more complex process than simple correlation may be required to account for human stereo processing and indicates other limitations of correlation models.