Dual Attention Message Passing Model for Scene Graph Generation
Zhendong Li, Gaoyun An, Songhe Feng, Qiuqi Ruan · 2019 IEEE 8th Data Driven Control and Learning Systems Conference (DDCLS) · 2019
Recently, scene graph has emerged as a powerful tool for enhancing the performance of scene understanding. Many algorithms have been proposed to generate scene graph and most of them include message passing process which is used to transfer information between nodes in scene graph. However, these algorithms did not fully consider the imbalance of information transmission in their message passing process. To address this defect, a novel Dual Attention Message Passing (DAMP) model consisted of internal and external attention mechanisms is proposed in this paper, which can regulate and control information transmission robustly. In our model, external attention mechanism is used to determine the importance of all nodes to the target node, and internal attention mechanism is used to reinforce correlated information between the source node and target node. On Visual Relationship Detection datasets, the proposed model can achieve comparable results with the state-of-the-arts.