GT-BoW is a geometry-aware bag-of-words image-retrieval system for ground-texture localization. It uses the fixed scale and consistent orientation of downward-facing imagery to reject visually similar but geometrically inconsistent matches, improving both global localization and loop-closure detection.
Key Contributions
- An approximate k-means vocabulary with soft assignment
- Scale-binned vocabularies and orientation-based geometric verification
- Separate high-accuracy and high-speed variants for global localization and loop closure
Performance Highlights
- Increased global-localization mean average precision from 0.026 to 0.559
- Detected nearly 3× as many loop closures as a standard DBoW baseline