The collective stereopsis algorithm described in [Marr:76a] was historically one of the first ``cooperative'' algorithms based on relaxation proposed for early vision.
The goal in stereopsis is to measure the difference in retinal position (disparity) of features of a scene observed with two eyes (or video cameras). This is achieved by placing a fiber of ``neurons'' (one for each disparity value) at each pixel position. Each neuron inhibits neurons of different disparities at the same location (because the disparity is unique) and excites neurons of the same disparity at near location (because disparity tends to vary smoothly). After, convergence the activation pattern corresponds to the disparity field defined above.
Figure 6.44: Collective Stereopsis: (top left) Definition for geometry of stereoscopic vision. (bottom left) Neural Network Activity (top three layers disparity d=0, 1, 2) corresponding to real world structure illustrated. (right) Results of iterations for d=0 and d=2 layers of neurons. d measures disparity value for pixels.
The parallel implementation is based on a straightforward domain decomposition and the results are illustrated in Figure 6.44. They show the initial state of disparity computation and the evolution in time of the different layers of disparity neurons. Details are described in [Battiti:88a].