Occupancy-Grounded Room Segmentation Improves 3D Scene Graphs

Hierarchical 3D scene graphs (3DSGs) help indoor robots organize geometric and semantic information across scales, but existing methods construct the critical room layer from different spatial substrates—place clusters, wall planes, or segmentation outputs—making room nodes inconsistent and hard to evaluate. In a new paper, researchers from an undisclosed institution (authors Carlos Cueto Zumaya, Iacopo Catalano, Jorge Peña-Queralta, Wallace Moreira Bessa) introduce an occupancy-grounded pipeline that anchors room nodes to tracked free-space regions derived from occupancy decomposition. This gives each room an explicit polygonal footprint, enabling direct geometric evaluation against annotated room instances.

Evaluated on 12 Matterport3D scenes, the method recovers substantially more room instances than Hydra (a representative place-connectivity baseline), but at the cost of lower precision. The results show that occupancy-grounded anchoring improves recall significantly, yet both methods struggle to produce wall-accurate room boundaries, which remains an open problem. The work provides a clear benchmark for future research and includes open-source code, offering a practical step toward more reliable spatial understanding for autonomous indoor navigation.