Scene Graph Generation Based on Node-Relation Context Module

For better understanding an image, the relationships between objects can provide valuable spatial information and semantic clues besides recognition of all objects. However, current scene graph generation methods don’t effectively exploit the latent visual information in relationships. To dig a better relationship hidden in visual content, we design a node-relation context module for scene graph generation. Firstly, GRU hidden states of the nodes and the edges are used to guide the attention of subject and object regions. Then, together with the hidden states, the attended visual features are fed into a fusion function, which can obtain the final relationship context. Experimental results manifest that our method is competitive with the current methods on Visual Genome dataset.