Competitive Hopfield neural network for stereo vision correspondence

Abstract

For the correspondence problem in stereo vision, we developed an Hopfield algorithm that favors unicity of the matches for all interest points both on the left and right images. Although the convergence of the network used cannot be theoretically proven, we have experimentally shown that for the cases we are interested in the method converges either to a stable state or to an acceptable limit cycle. The method is computationally fast.