acceptodds
Under review as a conference paper at ICLR 2027

Aqua3R: Feed-Forward Underwater Reconstruction with Dual Gaussian Scenes

Abstract

Water attenuation and backscatter couple scene appearance with the surrounding medium, so underwater reconstruction requires both reproducing the observed images and recovering the underlying water-free scene. Per-scene methods can address both goals, but require optimization from calibrated cameras, while existing feed-forward approaches do not explicitly separate in-water and restored scene representations. We present Aqua3R, a pose-free feed-forward model that predicts cameras, in-water and restored Gaussian scenes, and an explicit medium model in a single forward pass. The two scene representations share camera estimates and pre-fusion depth points and diverge only at the attribute heads. The medium model then re-renders the restored scene to reproduce the underwater observation. We further introduce DL3DV-Water, which provides paired clean and underwater views with formation parameters, and DeepSea-ROV, which contains raw deep-sea videos with registered clips. We also construct transfer sets and an evaluation protocol covering all predicted outputs. At 24 views, Aqua3R improves in-water PSNR by 0.98–2.75 dB over the stronger of two underwater-adapted YoNoSplat baselines on real footage and two synthetic transfer sets, while producing substantially more accurate camera estimates on real footage. In a controlled data study, broader medium coverage improves physical re-rendering PSNR on 25 real clips by an average of 1.06 dB. In contrast, gains from increasing the number of source scenes depend on the training design, and additional conditions or optimization updates do not yield consistent improvements. We will release the data and code, along with model checkpoints and per-scene results.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
underwater reconstruction, data scaling, feed-forward 3D reconstruction, 3D Gaussian splatting, novel view synthesis, image restoration, physics-based rendering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_dGruOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents Aqua3R, a feed-forward framework for underwater 3D reconstruction from unposed images. The method predicts camera parameters, an in-water Gaussian scene (W), a restored Gaussian scene (C), and scene-level medium parameters in one forward pass. The two Gaussian branches share camera estimates and pre-fusion depth-derived point support but predict separate Gaussian attributes. A physical image-formation model uses the predicted medium and rendered range to transform C into a water-affected rendering (P), with gradients stopped at C and its rendered depth. The authors also introduce the paired synthetic DL3DV-Water dataset and the real DeepSea-ROV collection. Experiments show improved in-water novel-view synthesis on real footage and synthetic transfer sets, together with improved camera estimation on Real25, relative to the evaluated baselines. The method trails YoNoSplat-D on the in-domain Test32 rendering benchmark. A separate data study investigates the effects of scene count, medium coverage, and training budget.

Critical strengths
  1. The paper addresses a practically important problem by jointly producing observed underwater and restored scene representations from unposed images without per-scene optimization. Predicting W and C under the same cameras is a useful design for obtaining directly comparable reconstructions and avoiding camera inconsistencies between separately trained models.

  2. The evaluation covers real underwater footage and two synthetic transfer sets, with ablations of the loss components and training-data design. Under the nominal 24-view protocol, Aqua3R improves in-water PSNR over the stronger adapted YoNoSplat baseline on each set from 18.55 to 19.53 dB on Real25 (Table 1, p. 6) and from 19.56 to 22.31 dB on OOD-Water (Table 8, p. 15). The additional training seed, paired confidence intervals, and scene-medium swap experiments provide useful context for interpreting these gains.

Critical weaknesses
  1. The experiments do not fully isolate the specific benefit of maintaining two Gaussian scenes. Section 3.1 (pp. 3-4) motivates the dual-scene design against an alternative that predicts only a restored scene and a medium, but no matched feed-forward comparison tests that alternative. Table 15 (p. 20) provides output-training controls with the physical and medium-label losses disabled. As explained in Appendix D.2 (p. 19), these controls freeze the branch not optimized by that control and mask its RGB loss while retaining both forward paths. They therefore do not directly test the complete dual-scene system against a clean-scene-plus-medium model. The shared-camera outputs offer a useful capability, but the performance benefit of the additional in-water representation remains uncertain without this matched comparison.

  2. The benefit of physical re-rendering over medium-label supervision alone remains untested. Table 6 (p. 9) removes the physical and medium-label losses together: Test32 C PSNR changes from 18.25 to 18.23 dB without a resolved difference, while Real25 P PSNR drops from 18.34 to 17.19 dB. Table 16 (p. 20) provides sensitivity results for nonzero physical-loss weights, and Table 17(a) demonstrates synthetic medium recovery. These results support the combined training objectives, but do not establish the incremental benefit of the physical loss over direct coefficient supervision. A control retaining medium-label supervision while disabling the physical loss would clarify the contribution of the physical consistency objective.

Questions for authors
  1. Could the authors compare Aqua3R with a feed-forward model that predicts only a clean Gaussian scene and a medium, using the same VGGT-Ω initialization, camera/depth supervision, training data, and optimization budget, while reporting model size? This would directly test the alternative discussed in Section 3.1 and clarify the benefit of the additional in-water representation.

  2. Could the authors add a control that retains medium-label supervision but sets the physical-loss weight to zero, with all other training settings fixed? Evaluating the resulting models on the same synthetic samples in the shared 4 m gauge, with C and P quality and coefficient errors reported together, would isolate the benefit of physical consistency beyond direct medium supervision.

  3. Appendix A.2 (p. 13) distinguishes possible upstream exposure of Test32 sources from the separate contamination concern associated with the original VGGT-Ω checkpoint. Could the authors assess whether the relative gains persist with the recommended retrained checkpoint, keeping the adaptation and evaluation protocol fixed, and distinguish the implications for Test32 from those for the transfer and real-data results?

Reviewer_sm3oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Aqua3R is a feed-forward model that takes unposed underwater images and, in one pass, predicts cameras, an in-water Gaussian scene (W), a restored water-free scene (C), and a scene-constant medium ω. Re-rendering C through ω gives P. Both scenes come from a VGGT-Ω backbone and share depth-derived Gaussian centres, splitting only at the attribute heads. The authors build DL3DV-Water (synthetic clean/underwater pairs rendered with the Akkaynak–Treibitz formation model) and DeepSea-ROV (registered real deep-sea clips), and run a data study on scene count, medium coverage and training budget.

Critical strengths

A practical problem with clear value: pose-free reconstruction and restoration of underwater scenes in about one second, versus minutes to hours for per-scene methods.

Clear gains in camera accuracy on real deep-sea footage, including all eight DeepSea-ROV clips.

Data contributions. DL3DV-Water (exact medium labels, same-camera clean references) and DeepSea-ROV (registered real clips split by dive) are useful if released.

Critical weaknesses

W1. The medium does not constrain restoration, and the stop-gradient that causes this was not validated where it can be measured. Requested experiments. Both can be run on DL3DV-Water, where clean references allow C to be measured directly:

  1. Detached vs. joint routing on synthetic data. Train (or adapt) the model twice under the same recipe, once with the stop-gradient and once with the physical loss allowed to reach C and its depth. Evaluate C against the clean references on Test32 and the transfer sets.
  2. Constrained routing. Instead of fully detaching C, test intermediate options that let the physical loss reach C. For example: a reduced weight on the physical gradient into C a regulariser that penalises drift of C from its unadapted prediction

Why this matters? These experiments would show whether the stop-gradient is a principled design choice or a workaround, and whether the paper’s central physics-based claim holds.

W2.The C branch is a copy from W and receives a skip connection from the raw underwater image. Both may bias C toward underwater colours. Please compare initialising C from AnySplat or randomly, and removing F_rgb from the C branch.

W3: The claim that W and C share one geometry is never checked. Only the depth rendered from C is supervised (L_d and L_dc). W shares the pre-fusion centres, but its opacity, scale and rotation are trained only to match colours. W could therefore imitate backscatter with semi-transparent Gaussians in front of the true surface, rendering a closer surface than C without any penalty. Experiment: render depth from W and from C at the same cameras on Test32 and Real25, and report their difference (e.g., mean absolute error and the percentage of pixels differing by more than 5%). Also report the effect of applying L_d to W’s rendered depth. Contradicting Example : A rock sits 4 m from the camera, with blue haze from the water in front of it. C must render only the rock, and its depth is supervised, so it must place the rock at 4 m. W must render the rock plus the haze, and it has two ways to do that: Correct: tint the rock itself blue. W’s depth stays at 4 m Shortcut: place semi-transparent blue Gaussians floating in the water, about 1 m from the camera. W’s depth becomes 1 m (contradictory)

Questions for authors

Questions :Is there any quantitative evidence for C on real water, for example colour-chart error on the SeaThru-NeRF scenes?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.