Recent years have seen a dramatic rise in the adoption of convolutional neural networks (ConvNets) for a myriad of computer vision tasks. The structure of convolution is proved to be powerful in numerous tasks to capture correlations and abstract conceptions out of image pixels. However, ConvNets are also shown to be deficient in modeling quite a few properties when computer vision works towards more difficult AI tasks. These properties include pairwise relation, global context and the ability to process irregular data beyond spatial grids.

An effective direction is to reorganize the data to be processed with graphs according to the task at hand, while constructing network modules that relate and propagate information across the visual elements within the graphs. We call these networks with such propagation modules as graph-structured networks. In this tutorial, we will introduce a series of effective graph-structured networks, including non-local neural networks, spatial generalized propagation networks, relation networks for objects and multi-agent behavior modeling, graph networks for videos and data of 3D domain. We will also discuss how to utilize graph-structured neural architectures to study the network connectivity patterns. Lastly, we will discuss related open challenges that still exist in many vision problems.