Disparity estimation from a stereo pairs using recurrent neural network

This paper presents a recurrent neural network based approach for disparity estimation from a stereo image pair. The network is trained to learn the uniqueness and continuity constraints using random dot stereo image pairs with the help of a new recurrent backpropagation algorithm proposed by us. However, in view of the large size of the network required we have implemented the algorithm on a piece-meal basis using small sized neural network after dividing the original image into parts. Further, the problem of large interconnection matrix was solved by taking the advantage of the sparseness of the weight matrix and uniformity of the network structure. Our experimental results shows that recurrent network is a viable alternative to Hopfield network for static stereo problems.