Kuan and Jaw ask how accurate the camera-to-scene distance has to be before Sea-thru colour restoration works. They build a millimetre-accurate scene in air (an 8 x 8 m patch with cartons and colour boards, 140 photos, 32 markers measured by total station, 6 mm RMSD at control points), simulate underwater versions for eight Jerlov water types at 1 m and 5 m depth, add zero-mean random error to the depth map (standard deviation 0.50, 0.30, 0.10, 0.05 and 0 m), run Sea-thru, and score against the in-air photo with a modified SSIM (luminance and contrast terms only, weighted by the eye’s luminous efficiency).
Findings
- In the clearest water (Jerlov type I, 1 m deep), one view scores 0.638, 0.790, 0.899, 0.931 and 0.933 as depth error falls from 0.50 m to zero. The authors conclude that about 0.30 m of depth accuracy is the floor for SSIM 0.8.
- In the more turbid types (IB, 1C) the 0.8 bar is mostly missed even with perfect depth: one view in type 1C scores 0.548 with zero error.
- Near regions (about 2 m away) follow the trend. Middle (about 4 m) and far (about 7 m) regions behave inconsistently, so Sea-thru is most trustworthy close up.
Limits
- One scene, and the water is simulated with the same physical model the restoration inverts (the authors note the two algorithms differ in detail). Real water adds uneven light and particles.
- The 0.8 SSIM bar comes from a survey of other papers, not from any downstream task.
- The grey-world step needs varied colours; the authors suspect their scene had too few at depth.
Relevance
A concrete number for how good depth must be before physics-based colour correction pays off. If a learned depth model on a small robot is 30 cm off at a few metres, the correction is already marginal, and in turbid water it fails anyway. See underwater perception and cai-2025-underwater-metric-depth.