A framework for refining object representations in semantic maps
This system uses probabilistic point-level confidence updating to improve 3D semantic maps.
Open-vocabulary 3D semantic maps serve as spatial memory for long-horizon robot autonomy. These maps allow robots to reason about objects and spatial relationships.
Existing object-centric semantic maps rely on an append-only point integration strategy. This strategy accumulates newly observed points but rarely reevaluates previously integrated points.
Perception errors accumulate over time. Outdated points persist after objects move or disappear.
The proposed online refinable semantic mapping framework continuously refines object representations. It uses probabilistic point-level confidence updating.
Each object point maintains a confidence score. This score is updated using multi-view temporal evidence like consistent observations and visibility cues.
The framework includes object tracking and re-identification. This preserves instance identities under motion and relocation. It enables maps to be continuously refined rather than monotonically accumulated.
Key facts
- Mapping accuracy improved by 3.46% on Replica and 2.55% on ScanNet.
- The framework shows success rates of 82.4% for object navigation and 76.7% for embodied question-answering.
- The system suppresses spurious and outdated points using inconsistent segmentation masks and depth measurements.