Through‑wave Imaging for littoral ISR by unsupervised learning of the water surface shape
- Passive “see-through” from the air: Recover an undistorted underwater view and the sea-surface height from a short RGB burst, no labels, calibration targets, or special optics. SIREN = sinusoidal representation network (an implicit neural field that models continuous signals and their derivatives).
- Better than the leading unsupervised baseline: Consistently outperforms NDIR (Non-rigid Distortion Removal) on real and simulated data, and competes with supervised single-image methods, while also estimating the surface itself.
- Operational fit for littorals: Enables standoff ISR from UAVs/USVs to aid MCM, route survey, harbor security, and object reacquisition; the recovered surface field provides sea-state cues for autonomy and sensor fusion.
- Clear limits, clear roadmap: Assumes nearly static scenes and near-orthographic views; ongoing work targets perspective-aware modeling, turbid-water robustness, and real-time edge deployment.
