Learning to transfer focus of graph neural network for scene graph parsing

Highlights：

• A new neural network architecture, the graphical focal network, is proposed to improve the recognition rate of semantic relationship in scene graph parsing task.

• The proposed graphical focal loss transfers the focus of network learning to the semantic relationship types with high value but limited instances.

• The proposed relative depth encoding module and regional layout encoding module introduce effective 3D spatial layout information.

• On the two evaluation metrics of scene graph parsing tasks, our method has achieved new advanced performance on the Visual Genome dataset.