Robustness of learned feature matchers under severe underwater turbidity

Determine whether learned image matchers can remain effective for structure-from-motion registration under severe underwater turbidity, particularly at turbidity levels where classical SIFT-based registration fails almost completely.

Background

The paper evaluates classical SIFT-based structure-from-motion registration on the SOTRUE dataset, whose sequences share identical trajectories while varying in measured turbidity. Registration succeeds for 99.5% of frames in clear water, falls to 1.0% at 7 NTU, and reaches 0.0% at 12 NTU. These results indicate a sharp failure of the conventional feature-matching pipeline under increased turbidity.

The authors explicitly delimit their finding to the SIFT-based pipeline used by the evaluated methods and do not test learned matchers. The unresolved issue is therefore whether learned matching methods can overcome the registration cliff in highly turbid underwater imagery.

References

These numbers characterise classical SIFT-based SfM, the pipeline the surveyed methods actually use; whether learned matchers survive the cliff is an open question this study does not test.

— Gaussian Splatting Underwater: A Controlled Cross-Regime Study  (2608.25483 - Álvarez-Tuñón et al., 26 Aug 2026) in Section E3, “Pose Source” (Sec. 4.3)

Demonstrating that cross-medium transfer, together with naturally drifting field channels, remains future work.

— Learning to deform the matched filter  (2608.31149 - Haigh, 31 Aug 2026) in Discussion, paragraph beginning “The study has several limitations”