This framework gives each node of an open-vocabulary 3D scene graph a probability of being correct, using only the detector confidence and embeddings the graph already stores, with no resampling, training or held-out labels. Four signals are mapped to probabilities at the VLM’s learned logit scale and propagated up the containment hierarchy into a belief that a room contains a queried class. Objects from the same detector and vocabulary are treated as failing together, so one repeated error is not counted as repeated evidence; on HM3DSem it improves object retrieval and room-level assertions over a strong 3DSG system.
Relevance
Counting correlated detections as independent is the same mistake as double-counting a sensor, and it is common in semantic mapping. Read for 3D scene graphs and robust estimation in multi-robot perception. The fully correlated view also applies to a team whose robots all run the same detector, since their agreement is weaker evidence than it looks.
Links
- arXiv
- Subfield: Semantic mapping, Robot learning