Scene graph representation robust against graph errors for scene graph classification

Abstract

In this research, we have developed a new scene graph representation for scene graph classifiers. Scene graphs are powerful scene representations that can represent scene structure (semantic and spatial structure). Recent progress in Graph Convolutional Network (GCN) has enabled training of scene graph classifier. However, the effective scene graph descriptors for GCN have not been established yet. In this paper, we developed a new method of scene graph representation using local features (PatchNetVLAD) that are robust against graph errors. We verified the effectiveness of the proposed method by experiments using the NCLT dataset.