I present an algorithm for calculating the likelihood of obtaining stereo images under the assumption that both eyes view the same scene, and test this algorithm using numerical simulations. Each stereo pair analyzed was based on two slightly different views of the same three-dimensional (3-D) scene (binocularly correlated) or two entirely different views of unrelated scenes (binocularly uncorrelated). The correlated stereo pair contained local binocular disparities that defined a consistent 3-D structure. A likelihood ratio was computed for each stereo pair, by assuming the increase of the local cross-correlation due to the epipolar constraint for binocular correspondence. For complex, random-dot and natural stereo images, the likelihood ratio correctly predicted whether the two eyes viewed the same 3-D scene. The results indicate that the algorithm proposed here is useful in modeling human stereopsis.
All Science Journal Classification (ASJC) codes