Wyner-ziv Coding of Stereo Images with Unsupervised Learning of Disparity

Wyner-Ziv coding can exploit the similarity of stereo images without communication among the cameras. For good compression performance, the disparity among the images should be known at the decoder. Since the Wyner-Ziv encoder has access only to one image, the disparity must be inferred from the compressed bitstream. We develop an Expectation Maximization algorithm to perform unsupervised learning of disparity at the decoder. Our experiments with natural stereo images show that the unsupervised disparity learning algorithm outperforms a system which does no disparity compensation. It is also superior to conventional compression using JPEG.