Visual enhancement and 3D representation for underwater scenes: a review

3 min read

Huang et al. review underwater image enhancement and underwater 3D reconstruction in one place. Their argument is that enhancement is mostly used as a pre-processing step for geometry, and its effect on that geometry is rarely measured. They start from the Jaffe-McGlamery image formation model (direct transmission, forward scatter, backscatter) and its simplified forms, survey enhancement from histogram and Retinex methods through physics-based, CNN, transformer, Mamba and diffusion models, then cover photogrammetry (SfM, visual SLAM, MVS), underwater NeRF and underwater 3D Gaussian Splatting. They close with case studies running COLMAP and Instant-NGP on raw and on enhanced underwater images.

Findings

  • In clear seawater, red light (600 to 700 nm) reaches less than 5 m and blue (400 to 500 nm) reaches 50 to 100 m, which is where the blue-green cast comes from. Backscatter usually does more harm than forward scatter because it adds a veil over the whole image.
  • They sort underwater NeRFs into four families: separating medium from object (SeaThru-NeRF), joint colour correction (WaterNeRF, UWNeRF), handling floaters and moving distractors (AquaNeRF), and restoration terms in the loss (WaterHE-NeRF). Underwater 3DGS splits the same way: medium-aware splatting (UW-GS, SeaSplat, WaterSplatting), depth or structure priors (Z-Splat, which fuses sonar with RGB), temporal degradation models (MarineSTD-GS), and restoration-coupled training (RecGS, R-Splatting).
  • On the SeaThru-NeRF dataset, using numbers reported by the original papers, average PSNR rises from 26.11 for SeaThru-NeRF and 26.72 for vanilla 3DGS to 30.91 for AtlantisGS.
  • Public underwater reconstruction datasets are small. SeaThru has 4 scenes and 88 images, S-UW 4 scenes and 96, Submerged3D 4 and 80; the largest has 23 scenes and over 2,000 images.
  • Enhancement can raise the number of detected features without raising the number of matches that survive geometric checks, and per-frame enhancement introduces colour differences between views. Their practical advice: use raw or lightly corrected images for pose estimation, and save strong enhancement for dense reconstruction or display.
  • Cost: NeRF training takes hours to days; 3DGS trains faster but its COLMAP initialisation still takes hours; diffusion enhancement is offline only; transformer enhancers need high-end GPUs for anything near online.
  • MarineInst, built on SAM and trained on MarineInst20M, is the only marine-specific foundation model they found. It does instance segmentation and captioning but was not trained to understand geometry.

Limits

  • Cameras throughout. Sonar gets one short section on multi-sensor systems, and the review deliberately leaves out detection, segmentation and tracking.
  • The comparison tables are compiled from the original papers without rerunning them under one protocol, which the authors say.
  • Most benchmarks report image metrics (PSNR, SSIM, LPIPS). The authors point out that Chamfer distance, ATE and RPE are rarely reported alongside them, so a method can look good on images while hurting localisation.
  • Static scenes dominate, refraction through flat ports is still unsolved for metric work, and stitching reconstructions from several vehicles with different lighting is listed as an open challenge.

Relevance

Any feature computed on underwater images inherits the colour shift and haze this review catalogues, and per-frame enhancement can change features between views of the same scene. The same holds for foundation models used in the water and for underwater perception generally.