Image-to-Image Retrieval by Learning Similarity between Scene Graphs

As a scene graph compactly summarizes the high-level content of an image in structured and symbolic manner, similarity between graphs two images reflects relevance their contents. Based on this idea, we propose novel approach for image-to-image retrieval using measured by neural networks. In our approach, networks are trained to predict proxy measure, computed from human-annotated captions pre-trained sentence model. We collect publish dataset human annotators evaluate algorithms. The collected shows that method agrees well with perception than other competitive baselines.