This paper introduces a novel contour-based approach named deep snake for real-time instance segmentation. Unlike some recent methods that directly regress the coordinates of the object boundary points from an image, deep snake uses a neural network to iteratively deform an initial contour to the object boundary, which implements the classic idea of snake algorithms with a learning-based approach. For structured feature learning on the contour, we propose to use circular convolution in deep snake, which better exploits the cycle-graph structure of a contour compared against generic graph convolution. Based on deep snake, we develop a two-stage pipeline for instance segmentation: initial contour proposal and contour deformation, which can handle errors in initial object localization. Experiments show that the proposed approach achieves state-of-the-art performances on the Cityscapes, Kins and Sbd datasets while being efficient for real-time instance segmentation, 32.3 fps for 512$\times$512 images on a 1080Ti GPU. The code will be available at https://github.com/zju3dv/snake/.