Scene Recognition Meets Knowledge Graphs: Enhancing Robustness to Object Diversity

Keisuke Yamashita, Seiya Ito, Tomoaki Yamazaki, Kouzou Ohara

Abstract: Scene recognition involves classifying an input image into one of a set of predefined scene classes. Accurate recognition requires understanding not only the objects present in the scene but also their spatial arrangements and the surrounding background. However, learning the associations between scenes and objects from image datasets can be challenging, especially when each scene class contains a wide variety of objects. To address this issue, we propose a scene recognition approach that leverages a commonsense knowledge graph to supplement the implicit relationships between scenes and objects that are difficult to infer directly from visual data alone. Experimental results demonstrate that incorporating the knowledge graph leads to improved scene recognition accuracy, validating the effectiveness of our proposed approach.

External IDs:dblp:conf/mva/YamashitaIYO25