AdaptVPR: Route-Aware Hard Positive Generation for Robust Visual Place Recognition
AdaptVPR builds 160K verified synthetic same-place positives to make visual place recognition less brittle under appearance shifts.
The method routes image edits through global weather/light/time changes, local dynamic occlusions, or both. A vision-language model estimates scene attributes and edit feasibility, while verification checks geometric consistency and appearance diversity. The authors report consistent gains across VPR baselines and vision foundation backbones, including R@1 improvements up to 9.2% under harder domain shifts. Source: HF Daily Papers' note
The method routes image edits through global weather/light/time changes, local dynamic occlusions, or both. A vision-language model estimates scene attributes and edit feasibility, while verification checks geometric consistency and appearance diversity. The authors report consistent gains across VPR baselines and vision foundation backbones, including R@1 improvements up to 9.2% under harder domain shifts. Source: HF Daily Papers' note
score 4