This paper releases a curated dataset for long-term visual localisation on the seafloor and a better way to define ground truth for it. The data are georeferenced AUV images from five benthic reference sites revisited over periods of up to six years, with raw and colour-corrected stereo imagery, calibrations and registered camera poses at sub-decimetre accuracy. Ground truth links two images when their estimated 3D seafloor footprints overlap, instead of when their camera positions are close.
Findings
- Eight visual place recognition methods score markedly lower Recall@K here than on established terrestrial and underwater benchmarks (abstract).
- Distance-threshold ground truth overestimates Recall@K at sites with rugged terrain and changing altitude, compared with footprint-based ground truth.
Limits
- Abstract only so far. Near-nadir AUV imagery; forward-looking diver or small-ROV views are a different problem.
Relevance
It tests whether an underwater robot can relocalise against imagery taken years earlier, and warns that the usual “within X metres” ground truth flatters methods on rough terrain. Pairs with boittiaux-2023-eiffel-tower (deep sea, one site). See underwater perception.
Links
- arXiv
- Subfield: Field robotics, Semantic mapping