William Freeman Massachusetts Institute of Technology Many vision problems involve gathering and propagating local evidence. A convenient framework for that, when the image interpretation is structured as a probabilistic graphical model, is belief propagation. When the graphical model is a chain or a tree, this calculation is exact. When the graphical model has loops, the belief propagation algorithm is only approximate. I'll show applications of the belief propagation algorithm for data-driven low-level vision problems, posed as interpreting local image patches. One example is image super-resolution (estimating a high-resolution image from a low-resolution input); another is the analysis of image motion.