Geometry beneath the Waves: Dense Priors for Sparse-View Underwater 3D Gaussian Splatting
Harvey Caldeira, Haoran Wang, Guoxi Huang, Shaoyu Cai, Rachel Fu, Nantheera Anantrasirichai
cs.CV
Quick verdict
Underwater-adapted VGGT plus 3D-consistent pseudo-views boost sparse-view 3DGS by 2.74 dB on SeaThru-NeRF but give near-zero, SSIM-negative gains on Submerged3D.
Summary
The paper tackles the geometry-initialization bottleneck of sparse-view underwater 3D Gaussian Splatting (3DGS). It argues that water degradation breaks COLMAP feature matching, while feed-forward geometry models such as VGGT applied zero-shot underwater produce volumetric water-column noise and unstable surfaces. Two contributions follow. First, domain-adapted feed-forward geometry: keeping clean geometric labels and synthetically degrading images with SyreaNet's physics-based underwater model over 22,000 images (1,000 Hypersim scenes, 1,000 TartanAir images, 20,000 SA-1B images with Depth Anything 3 pseudo-depth), VGGT is fine-tuned through Fin3R teacher-student distillation with LoRA adapters confined to the visual aggregator. Second, geometry-guided view synthesis: the adapted VGGT initializes a 3DGS proxy that renders pseudo-views with pseudo-depth along the camera trajectory; COLMAP re-runs jointly over captured and rendered views to expand feature tracks; RUSplatting performs final optimization and color restoration. On SeaThru-NeRF and Submerged3D (four scenes each, ~20 views, 720p), the pipeline reports 27.11 dB PSNR on SeaThru-NeRF, +2.74 dB over COLMAP-initialized RUSplatting, but on Submerged3D only 23.49 vs 23.48 dB with SSIM below the COLMAP-to-3DGS baseline. Stated limitations include COLMAP failure on IUI3-RedSea and residual prediction noise; no code release is mentioned.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Reversed supervision - preserving clean geometry while synthetically degrading images with a physics-based underwater model - adapts VGGT to underwater imagery more reliably than post-hoc alignment of noisy real-image depth.
- Rendering pseudo-views from an explicit 3DGS proxy and re-running COLMAP over captured plus rendered views expands feature tracks more effectively than RUSplatting's 2D frame interpolation.
- The full pipeline improves SeaThru-NeRF from 24.37 to 27.11 dB PSNR over COLMAP-initialized RUSplatting and reports the best PSNR (23.49 dB) and LPIPS (0.3322) on Submerged3D.
- LoRA adapters restricted to VGGT's visual aggregator preserve pretrained metric-scale geometric priors while enabling encoding of degraded underwater images.
Strengths
Clear failure-mode diagnosis motivating the design
Section 1 identifies two specific artifacts of zero-shot feed-forward geometry underwater - volumetric water-column noise and high-frequency surface dispersion - and links RUSplatting's 2D frame interpolation to geometrically inconsistent training views, giving the method a concrete rationale rather than generic motivation.
Reversed supervision avoids noisy underwater pseudo-labels
Section 2.1 keeps clean geometry and degrades images instead, sampling attenuation, backscatter, and transmission per frame via SyreaNet's physics-based model; this sidesteps the acknowledged problem that real-image pseudo-depth inherits the very scattering artifacts the model must remove.
Parameter-efficient adaptation preserves pretrained priors
The initializer grid isolates component contributions
Weaknesses
Submerged3D gains are effectively a tie and SSIM regresses
No variance or per-scene reporting
Missing feed-forward baselines and ablations
Synthetic-only training data with domain gap
The pipeline still depends on COLMAP mid-stage
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What problem does this paper solve?
It improves 3D reconstruction of underwater scenes from a small number of photos. In turbid water, the standard COLMAP pipeline fails because particles and wavelength-dependent absorption break feature matching, while off-the-shelf AI geometry models like VGGT hallucinate spurious 'ghost' geometry in open water. The paper adapts VGGT to underwater imagery using synthetically degraded training data and then uses its predicted 3D geometry to render extra, geometrically consistent training views for 3D Gaussian Splatting.
How much does the method improve underwater reconstruction?
On the SeaThru-NeRF benchmark the full pipeline reports 27.11 dB PSNR versus 24.37 dB for the strongest prior baseline (COLMAP-initialized RUSplatting), a +2.74 dB gain with better SSIM and LPIPS. On the Submerged3D benchmark the improvement is essentially nil: 23.49 vs 23.48 dB PSNR, with structural similarity (SSIM 0.7045) actually below the COLMAP-to-3DGS baseline (0.7385).
Is code or data available for this paper?
The paper text contains no code or data release statement. Training and evaluation details are described (22,000 synthetically degraded training images from Hypersim, TartanAir, and SA-1B; 720p evaluation; specific iteration counts and thresholds), but LoRA rank and optimizer settings are not specified, so reproduction currently requires re-deriving several choices.
What are the main limitations of this approach?
The authors state two: the mid-pipeline COLMAP stage fails entirely on the IUI3-RedSea dataset because scattering disrupts SIFT matching, and low-amplitude prediction noise persists under sparse capture. Reviewers would add: gains on Submerged3D are within noise, results are averaged over only four scenes per dataset without variance, and competing feed-forward models such as DUSt3R/MASt3R were not compared.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1 and 2 could not be inspected from the extracted text, so qualitative claims about suppressed water-column noise and coherent depth rest on the authors' descriptions.
- Only dataset-level averages appear in the text; per-scene breakdowns are unavailable, limiting assessment of consistency across scenes.
- This is a SIGGRAPH Asia Posters-track paper of roughly two pages, so the absence of extensive ablations and statistical detail partly reflects the format rather than negligence alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.