The Eiffel Tower dataset benchmarks long-term visual localisation in the deep sea. It holds images from four visits to the same hydrothermal vent edifice over five years. Camera poses and a shared scene geometry were estimated from vehicle navigation data and structure from motion, giving a reference to score localisation methods against.
Findings
- The authors analyse what changed between visits; underwater, the sources of variation are water conditions and the growth of marine organisms (abstract).
- Several established visual localisation methods leave clear room for improvement on it.
- Data are public on SEANOE.
Limits
- Abstract only so far. One site, and deep-sea lighting comes entirely from the vehicle.
Relevance
The data show slow, biological change that a robot relocalising against old imagery has to absorb. Pairs with larsen-2026-benthic-localization for shallower, multi-site seafloor data. See underwater perception.
Links
- arXiv · data
- Subfield: Field robotics, Semantic mapping