Persistent Metric-Semantic Maritime Scene Memory from Open-Vocabulary 3D Scene Graphs

This research investigates the transfer of outdoor open-vocabulary 3D scene graph methods to maritime environments. The authors demonstrate that while zero-shot descriptions agree with ground truth for many objects, existing maps suffer from fragmentation and fail to capture maritime spatial organization. To address this, they introduce a metric-semantic consolidation method combined with geometry-grounded multi-view vision-language model reasoning to reconcile fragments into physical entities and group them into local places. Additionally, macro regions are constructed using map priors and onboard observations. The resulting representation reduces fragmentation, supports correspondence of fixed local places across repeated traversals, and improves contextual retrieval compared to tested baselines.