Image-to-image scene matching by representing semantic structures using scene graphs

We propose a scene matching method that is robust to viewpoint changes and the presence of similar objects by leveraging scene-graph features that capture the semantic structure of the environment. We utilize the Weisfeiler-Lehman kernel to extract these features, capturing the topological structure of the scene graphs for similarity comparison. We constructed a dataset comprising 10 domestic scenes, each with reference images defined as user-specified goal locations. For each scene, similarities between the reference and query images captured from different viewpoints were computed. Experimental results show that our method accurately matches scenes even in the presence of similar objects and viewpoint variations.