Contextual Graph Reasoning Networks
Graph Reasoning has shown great potential recently in modeling long-range dependencies, which are crucial for various computer vision tasks. However, the graph representation learned by existing methods is not effective enough as the relation between feature and graph is under-explored. In this work, we propose a novel method named Contextual Graph Reasoning (CGR) that learns a context-aware relation between feature and graph. This is achieved by constructing the projection matrix based on a global set of descriptors during graph projection, and calibrating the evolved graph based on the self-attention of all nodes during graph reprojection. Therefore, contextual information is well explored in both graph projection and reprojection with our method. To verify the effectiveness of our method, we conduct extensive experiments on semantic segmentation, instance segmentation, and 2D human pose estimation. Our method consistently achieves remarkable improvements over state-of-the-art methods, demonstrating the effectiveness and generalization ability of our method.
PDF Abstract
MS COCO
ADE20K
COCO-Stuff
