Organizations: College of Computer Science, Nankai University · NKIARI, Shenzhen Futian · School of Automation, Southeast University · Advanced Ocean Institute of Southeast University, Nantong
Underwater 3D reconstruction faithfully reproduces the color shifts and visibility loss of captured views, while physical inversion may leave estimation errors in the recovered scene appearance. We formulate Underwater Scene-level Enhancement (USE) as learning a persistent, visibility-enhanced 3D scene representation from degraded multi-view underwater observations, enabling consistent enhanced rendering. Realizing USE requires both a reliable scene representation for enhancement and a consistent enhancement target without paired enhanced 3D data. Therefore, we present 3D-USE, a two-stage framework. First, the Medium Radial Basis Anchor Representation (MediumRBF) establishes a medium-aware Gaussian scene by representing water effects with shared radial-basis anchors and explicitly decomposing object and medium contributions. Based on this fixed scene representation, Appearance Transition Consensus (ATC) transfers paired 2D underwater image enhancement (UIE) knowledge into scene-global and Gaussian-local targets, avoiding direct supervision from inconsistent enhanced views. An Underwater Bilateral Appearance Field (U-BAF) then realizes these targets in Gaussian radiance and medium appearance. The scene directly renders enhanced novel views without a 2D UIE model at inference. Experiments on real underwater scenes show improved visibility and cross-view consistency while preserving reconstruction quality.
Figures & tables
Figure 1: Why scene-level underwater enhancement is necessary. Physical inversion, independent view enhancement, and the absence of enhanced 3D supervision prevent a stable shared scene appearance (a-c). 3D-USE instead transfers 2D enhancement knowledge into a persistent 3D representation for consistent enhanced novel-view rendering (d).
Figure 2: Overview of 3D-USE. Stage 1 reconstructs a medium-aware Gaussian scene with MediumRBF. An offline paired-data calibrator maps captured-view UIE proposals into a common transition space, where ATC forms fixed scene-global and Gaussian-local targets. With Stage 1 frozen, Stage 2 optimizes U-BAF through the transition realized by the complete enhanced render, storing the result in object and medium appearance for novel-view rendering without a 2D UIE model.
Figure 3: Qualitative comparison of underwater scene reconstruction. Zoomed regions highlight background consistency and medium-related artifacts. 3D-USE produces cleaner scene appearance with fewer medium-induced inconsistencies.
SeaThru-4
DRUVA-20
D3
D5
Method
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
FPS
WaterSplatting (3DV, 2025 )
29.340
0.919
0.129
27.007
0.864
0.357
27.380
0.776
0.253
29.113
0.817
0.313
99.78
SeaSplat (ICRA, 2025 )
27.146
0.878
0.184
22.866
0.828
0.388
27.885
0.785
0.282
27.113
0.806
0.382
100.77
MarineSTD-GS (ACM MM, 2025 )
30.034
0.901
0.160
26.472
0.856
0.364
29.510
0.781
0.316
30.237
0.818
0.329
41.13
Plenodium (NeurIPS, 2025 )
30.331
0.922
0.125
27.413
0.866
0.335
28.686
0.803
0.218
29.088
0.817
0.314
91.60
3D-UIR (IEEE TIP, 2026 )
28.260
0.889
0.179
25.432
0.820
0.354
28.098
0.830
0.283
28.973
0.823
0.324
42.74
Table 1: Quantitative comparison of reconstruction performance in terms of PSNR ↑ , SSIM ↑ , and LPIPS ↓ . SeaThru-4 and DRUVA-20 report averages across scenes. Best and second-best results are bold and underlined.
Figure 4: Qualitative comparison of enhanced novel-view renderings across different underwater scenes. Compared with existing methods, 3D-USE achieves more balanced color correction and visibility enhancement with fewer spatial artifacts.
SeaThru-4
DRUVA-20
D3
D5
Method
UCIQE
URanker
MUSIQ
UCIQE
URanker
MUSIQ
UCIQE
URanker
MUSIQ
UCIQE
URanker
MUSIQ
wLPIPS
WaterSplatting
0.508
1.361
58.529
0.385
− 0.856
33.193
0.470
0.541
58.833
0.499
0.690
55.568
0.173
SeaSplat
0.588
1.555
54.470
0.518
− 0.186
28.252
0.504
1.007
58.251
0.442
− 0.895
40.648
0.209
Plenodium
0.503
1.058
58.370
0.395
− 0.792
34.786
0.483
0.705
60.158
0.443
0.033
52.456
0.158
MarineSTD-GS
0.522
0.918
55.890
0.447
− 0.567
39.070
0.443
0.416
56.246
0.413
− 0.117
54.424
0.216
3D-UIR
0.609
1.658
54.540
0.537
0.174
32.752
0.519
1.356
59.351
0.468
0.137
51.991
0.208
Table 2: Quantitative comparison of enhancement quality and cross-view appearance consistency using UCIQE ↑ , URanker ↑ , MUSIQ ↑ , and wLPIPS ↓ . SeaThru-4 and DRUVA-20 report scene-macro averages, while wLPIPS is the scene-macro average over all 26 scenes. Best and second-best results are bold and underlined, respectively.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Vanilla 3DGS
27.411
0.889
0.178
MLP medium
29.340
0.919
0.129
MediumRBF w/o depth
30.212
0.922
0.125
Full Model (Ours)
30.369
0.924
0.124
Table 3: Stage-1 ablation on SeaThru-4. Best and second-best results are bold and underlined.
Figure 5: Qualitative Stage-2 ablation results. Arrows highlight inconsistent enhancement in the global-only variant. The full model produces more uniform and natural results.
Method
UCIQE ↑
URanker ↑
MUSIQ ↑
Stage 1
0.518
0.784
57.500
Local only
0.562
1.394
57.831
Global only
0.614
1.491
58.402
Full Model (Ours)
0.609
1.679
59.881
Table 4: Stage-2 ablation on SeaThru-4. Best and second-best results are bold and underlined.
Figure 6: Comparison of different strategies for integrating 2D UIE priors.
IUI3
Curasao
Japanese Gardens
Panama
Method
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
PSNR
SSIM
LPIPS
WaterSplatting
29.330
0.891
0.206
31.701
0.951
0.114
24.732
0.892
0.120
31.598
0.941
0.076
SeaSplat
27.552
0.853
0.209
29.323
0.892
0.191
22.777
0.867
0.190
28.932
0.899
0.146
MarineSTD-GS
30.433
0.890
0.202
33.688
0.939
0.143
26.337
0.881
0.152
29.678
0.893
0.145
Plenodium
30.179
0.892
0.205
33.861
0.955
0.105
24.872
0.895
0.115
32.413
0.945
0.075
3D-UIR
28.640
0.868
0.233
30.722
0.896
0.178
23.592
0.884
0.162
30.085
0.910
0.142
Table 5: Scene-wise reconstruction results on SeaThru-4 held-out views. Best and second-best results are bold and underlined.
Figure 7: Additional qualitative comparison of underwater novel-view reconstruction. Columns correspond to Japan (Japanese Gardens), A1, A2, A3, D3, and D5. The first row shows the held-out underwater views, followed by the reconstruction results of WaterSplatting, SeaSplat, MarineSTD-GS, Plenodium, 3D-UIR, and 3D-USE.
IUI3
Curasao
Japanese Gardens
Panama
Method
UCIQE
URanker
MUSIQ
UCIQE
URanker
MUSIQ
UCIQE
URanker
MUSIQ
UCIQE
URanker
MUSIQ
WaterSplatting
0.509
1.045
57.549
0.547
1.924
66.755
0.497
1.555
56.981
0.480
0.920
52.831
SeaSplat
0.579
1.503
51.292
0.586
1.247
60.171
0.601
1.922
53.844
0.586
1.549
52.572
Plenodium
0.511
0.826
60.681
0.499
0.831
61.631
0.507
1.489
57.784
0.496
1.086
53.384
MarineSTD-GS
0.546
1.161
58.370
0.477
0.087
58.286
0.543
1.329
55.681
0.520
1.094
51.222
3D-UIR
0.597
1.558
49.705
0.593
1.258
62.398
0.621
1.925
52.173
0.623
1.890
53.886
Table 6: Scene-wise no-reference enhancement quality on SeaThru-4 held-out views. Higher is better for all metrics. Best and second-best results are bold and underlined.
Figure 8: Additional qualitative comparison of enhanced novel-view renderings. Columns correspond to IUI3, Curasao, A2, A16, D3, and D5. Rows show the corresponding underwater inputs and the results of WaterSplatting, SeaSplat, MarineSTD-GS, Plenodium, 3D-UIR, and 3D-USE.
wLPIPS ↓
Method
SeaThru-4
DRUVA-20
D3
D5
WaterSplatting
0.1565
0.1699
0.2210
0.2453
SeaSplat
0.1530
0.2189
0.2131
0.2341
Plenodium
0.1446
0.1537
0.2124
0.2398
MarineSTD-GS
0.1447
0.2309
0.2108
0.2169
3D-UIR
0.1558
0.2174
0.2376
0.2017
Table 7: Dataset-wise cross-view consistency measured by wLPIPS (lower is better). Best and second-best results are bold and underlined.