Refractive COLMAP builds refraction into every stage of COLMAP’s structure-from-motion pipeline, for both flat-port and dome-port underwater housings. The usual practice is to absorb refraction into a pinhole camera’s distortion parameters, which is an approximation; this paper models the glass and water interfaces explicitly.
Findings
- In simulation and on photo-realistic synthetic images with ground truth, enabling refraction loses no accuracy or robustness compared with in-air reconstruction (abstract).
- Demonstrated at scale on a refractive dataset of nearly 6,000 images. Open source.
Limits
- Abstract only so far. Needs the housing geometry (port type and distances) to be calibrated.
Relevance
Every feed-forward model in this cluster assumes a pinhole camera, and a flat port breaks that assumption in a way that grows towards the image edges. Refraction-aware poses would also make a fairer reference for benchmarks like Water3D. See underwater perception.
Links
- arXiv · code
- Subfield: Field robotics, Mapping and localisation