Multimodal deep fusion for image question answering

摘要

Multimodal fusion plays a key role in Image Question Answering (IQA). However, most of the current algorithms are insufficient to fuse multiple relations implied in multimodalities which are vital for predicting correct answers. In this paper, we design an effective Multimodal Deep Fusion Network (MDFNet) to achieve fine-grained multimodal fusion. Specifically, we propose Graph Reasoning and Fusion Layer (GRFL) to reason complex spatial and semantic relations between visual objects and fuse these two kinds of relations adaptively. This fusion strategy allows different relations make different contribution guided by the reasoning step. Then a Multimodal Deep Fusion Network is built based on stacking several GRFLs, to achieve sufficient multimodal fusion. Quantitative and qualitative experiments conducted on popular benchmarks including VQA v2 and GQA reveal the effectiveness of DMFNet. Our best single model achieves 71.19% overall accuracy on VQA v2 dataset, and 57.05% accuracy on GQA dataset.