Visual Relation Extraction via Multi-modal Translation Embedding Based Model

Visual relation, such as "person holds dog" is an effective semantic unit for image understanding, as well as a bridge to connect computer vision and natural language. Recent work has been proposed to extract the object features in the image with the aid of respective textual description. However, very little work has been done to combine the multi-modal information to model the subject-predicate-object relation triplets to obtain deeper scene understanding. In this paper, we propose a novel visual relation extraction model named Multi-modal Translation Embedding Based Model to integrate the visual information and respective textual knowledge base. For that, our proposed model places objects of the image as well as their semantic relationships in two different low-dimensional spaces where the relation can be modeled as a simple translation vector to connect the entity descriptions in the knowledge graph. Moreover, we also propose a visual phrase learning method to capture the interactions between objects of the image to enhance the performance of visual relation extraction. Experiments are conducted on two real world datasets, which show that our proposed model can benefit from incorporating the language information into the relation embeddings and provide significant improvement compared to the state-of-the-art methods.