SceneGraphFusion: Incremental 3D Scene Graph Prediction From RGB-D Sequences

Figure 1 . We create a globally consistent 3D scene graph b) by fusing predictions of a graph neural network (GNN) from an incremental geometric segmentation created from an RGB-D sequence a). Our method merges nodes on the same object instance and naturally grows and improves over time when new segments and surfaces are discovered, see c). As a by-product, our method produces accurate panoptic segmentation of large-scale 3D scans. The nodes represent the different object segments.