Multiple Hub-Driven Attention Graph Network for Scene Graph Generation
Yao Yang, Bo Gu · 2021
Existing works of scene graph generation usually learn global context information leading to excessive redundant information being considered. As a matter of fact, most relationships only exist around several hub objects. To exploit this fact, we propose a novel Multiple Hub-driven Attention Graph Network (MHAGN) for scene graph generation. Specifically, we first classify objects into multiple subgraphs according to the degree of correlations with the hub objects. Then, we design Multiple Hub-driven Attention (MHA) to drive context information to be passed within multiple subgraphs separately, and force objects to attend more to associated objects in the sub-graph. Finally, MHAGN captures precise and diverse context information by combining MHAs from multiple subgraphs and generates compact relation-aware representations for objects. Experimental results on popular benchmarks show that the proposed MHAGN achieves better performance over baselines on several datasets, especially in terms of alleviating the imbalance of predicted relationship categories.