OceanXL: Large-scale Underwater 3D Gaussian Splatting via Block Partitioning and Adaptive Pruning

1 Introduction

• Underwater image degradation and uneven lighting can yield incomplete geometry, noisy reconstructions, and inconsistent appearance across viewpoints, especially in large scenes.The introduction identifies light absorption, scattering, wavelength-dependent color attenuation, and non-uniform illumination as sources of difficulty.Source: §1 · p. 2
• OceanXL partitions underwater scenes into spatial blocks, reconstructs them independently, and merges them into a coherent global representation.This divide-and-conquer design is intended to make optimization efficient for large scenes.Source: §1 · p. 2
• The introduced datasets span shipwrecks, coral reefs, and deep-sea environments, addressing the scarcity of large-scale underwater reconstruction datasets.The datasets are intended to support systematic evaluation of scalable reconstruction methods.Source: §1 · pp. 2–3
• Its underwater-aware density control compensates for attenuation-induced densification failures while progressively pruning redundant primitives.The stated aim is to improve scalability without sacrificing reconstruction fidelity.Source: §1 · p. 2
• Selective Difference-of-Gaussians activation adds representational capacity only in reconstruction-sensitive regions for compact, high-fidelity scene representations.The contribution describes this as a way to enhance local representational capacity.Source: §1 · p. 3

2.1 Underwater 3DGS Reconstruction

• Underwater 3DGS methods model underwater image formation or physical effects to improve reconstruction under scattering, attenuation, and visibility degradation, including UW-GS and WaterSplatting [Li et al. 2025].Other approaches use implicit water representations, radiance-medium disentanglement, recurrent refinement, uncertainty modeling, or image-enhancement priors.Source: §2.1 · p. 3

2.2 Pruning Gaussian Splats

• Gaussian pruning methods rank primitives using opacity [Navaneet et al. 2024; Zhang et al. 2025a], ray contributions, blending weights, spatial sensitivity, or gradients.Their effectiveness under severe underwater degradation and spatially varying visibility remains largely unexplored.Source: §2.2 · p. 3

2.3 3DGS-based Large Scene Reconstruction

• Large-scene 3DGS methods scale reconstruction with spatial partitioning, structured levels of detail, or system optimizations such as selective offloading and deferred updates.Examples include block-wise training in CityGaussian [Liu et al. 2024b], visibility-aware optimization in BlockGaussian [Wu et al. 2025], and host-GPU transfers in GS-Scale.Source: §2.3 · p. 3

3 OceanXL: Fast Large-scale Underwater Representation

• OceanXL reconstructs scenes by partitioning them into manageable subproblems, optimizing each independently, and merging the chunks into a unified representation.Source: §3 · p. 3

3.1 Balanced Scene Partitioning

• Fixed-size grids can create uneven point and camera distributions, so Balanced Scene Partitioning adapts splits to balance child point counts and computational load.It selects 2D or 3D partitioning based on scene geometry, and uses guard rails against partitions with insufficient geometry or camera coverage.Source: §3.1 · pp. 3–4
• Camera visibility determines chunk assignment; adding observed points across chunk boundaries supports consistency, while chunks with too few points or cameras are discarded.Source: §3.1 · p. 4

3.2 Underwater-Pruning Accelerated Gaussian Splatting

• Underwater-compensated Densification addresses attenuation-suppressed gradients with learnable depth-aware weights for mid- and far-range Gaussians.Near-region attenuation is assumed negligible; the method avoids an auxiliary network.Source: §3.2 · pp. 4–5
• Wise Adaptive Pruning progressively removes redundant or low-contribution primitives to reduce storage and rendering cost, while DoG components help preserve fine detail.The pseudo-Gaussian differs from the primary Gaussian in opacity and scale, increasing the expressive capacity of remaining primitives after pruning.Source: §3.2 · p. 5
• DoG activation is selective: reconstruction-sensitive primitives activate it when their unweighted 2D positional-gradient magnitude exceeds a threshold.This avoids unnecessary computation from global activation while retaining capacity in detail-critical regions.Source: §3.2 · p. 5

• Abyssal provides two wreck sequences, while OceanXplore provides three large-scale coral-reef video sequences.Source: §4 · pp. 5–6
• The datasets include varied capture conditions and scene content, including low-light distortion, dynamic marine life, illumination variation, and complex sea-fan structures.Source: §4 · p. 6
• The current OceanXplore release targets scalability rather than exhaustive coverage; its three sequences were selected for spatial extent and predominantly static structure.Transient fish motion remains, particularly in Komodo.Source: §4 · p. 6
• Evaluation also uses the public Eiffel Tower dataset and Tabuhan P1, a subset of Sweet Corals.Source: §4 · p. 6

• The large-scale comparison includes Hierarchical-GS, CityGaussian, Octree-GS, and BlockGaussian.Source: §5.1 · p. 6
• UW-GS combined with the proposed BSP strategy serves as an additional underwater-specific baseline.Source: §5.1 · p. 6
• Evaluation uses PSNR, SSIM, and LPIPS alongside model size and training time to assess scalability.Source: §5.1 · p. 6

• COLMAP initializes the point cloud and estimates camera poses; partitioning dimensionality is selected automatically.Source: §5.2 · p. 6
• ISRO and Tabuhan P1 use 2D ground-plane partitioning, while KWAJ and Eiffel Tower 2016 use 3D longest-axis splits for stronger vertical structure.Source: §5.2 · p. 6
• Each chunk is optimized for 40K iterations, with densification for the first 20K and modified pruning plus 3D-DoG for another 10,000 iterations; the final model retains 50% of original primitives.The reported training setup uses one RTX3090 GPU.Source: §5.2 · p. 6

5.3 Results and Comparison

• With a 40,000-point budget per partition, BSP uses unaligned boundaries to balance point-cloud distribution rather than splitting the scene at the midpoint.The comparison is on the ISRO scene against BlockGaussian partitioning.Source: §5.3.1 · p. 6

5.3.1 Partitioning Results.

• OceanXL uses only 32.3% as many primitives as BlockGaussian on average, with comparable-or-better PSNR, SSIM, and LPIPS overall.On Eiffel Tower 2016, Octree-GS is smaller and faster but OceanXL has higher reconstruction quality; on Tabuhan P1, OceanXL is larger than CityGaussian under the stated single-GPU constraint.Source: §5.3.2 · p. 6

5.3.3 Qualitative Comparison.

• On Eiffel Tower 2016 and Tabuhan P1, OceanXL recovers fine underwater structures while maintaining a compact scene representation.The passage also points to further qualitative comparisons and point-cloud visualizations.Source: §5.3.3 · p. 6

5.4 Ablation Study

• Replacing BlockGaussian partitioning with BSP improves reconstruction quality and training efficiency by balancing geometry across partitions and reducing reconstruction chunks.The passage attributes the efficiency gain to improved single-GPU optimization.Source: §5.4 · pp. 6–7
• The supplied Table 1 caption describes quantitative evaluation across five large-scale scenes, with model size in GB and OOM indicating out-of-memory failure.It identifies UW-GS* as a modified version combined with BSP.Source: Table 1 · p. 7
• Figure 3 compares OceanXL qualitatively with baseline approaches on Eiffel Tower 2016 and Tabuhan P1.Source: Figure 3 · p. 7
• Selective DoG activation with WAP gives a more favorable quality–compactness–efficiency trade-off than full activation, though its reconstruction metrics are slightly lower than V2 and V4.WAP without DoG reduces model size and training time but degrades quality; full DoG activation recovers quality with computational overhead.Source: §5.4 · p. 7
• UCD further improves PSNR, SSIM, and LPIPS by compensating for attenuation-induced densification failures, while increasing model size.This is the V2 variant in the ablation sequence.Source: §5.4 · p. 7

5.5 Generalization to Small-Scale Underwater Scenes

• 0.125 GB of storage gives OceanXL competitive SeaThru-NeRF reconstruction quality, with an 82–97% size reduction versus the evaluated underwater-specific baselines.The passage frames this comparison as evidence of compactness and small-scale generalization, not large-scale scalability.Source: §5.5 · p. 7
• The SeaThru-NeRF comparison reports PSNR, SSIM, LPIPS, and model size under the standard evaluation protocol.The Table 3 caption specifies that model size is reported in GB and indicates metric direction with arrows.Source: §5.5 · p. 7, Table 3 · p. 7
• A separate Table 2 caption describes average PSNR, SSIM, and LPIPS for variants on the Abyssal dataset with 3DGS as backbone.It distinguishes full from selective DoG activation and defines 3DGS* as using BlockGaussian partitioning.Source: Table 2 · p. 7

6 Discussion

• OceanXL’s results suggest scalability depends less on raw Gaussian expressivity than on how computation, capacity, and visibility are distributed across the scene.Source: §6 · p. 7

6.1 Effectiveness for Large-Scale Data

• BSP balances partition complexity to improve optimization stability, while adaptive pruning removes redundant densification primitives and selectively activates DoG in reconstruction-sensitive regions.The discussion links imbalance to costly dense partitions and unstable sparse ones, and notes underwater scattering and attenuation exacerbate redundant allocation.Source: §6.1 · pp. 7–8

6.2 Underwater Reconstruction Quality

• UCD uses three depth-dependent densification weights and improves underwater color consistency against standard large-scene GS, but its fixed depth boundaries may miss spatially varying optical properties.The discussion says more adaptive depth stratification could further improve reconstruction quality.Source: §6.2 · p. 8

6.3 Selective Activation of DoG Primitives

• Selective DoG allocates extra capacity to reconstruction-sensitive regions, avoiding overhead in low-detail areas and requiring no auxiliary neural decoding.The discussion notes compatibility with real-time rasterization and chunk-based VR/AR streaming, but says manually defined activation thresholds may not match perceptual importance across degradations.Source: §6.3 · p. 8

6.4 Failure Cases

• Sparse views can leave too little representational capacity after pruning to recover fine structures, while dynamic fish may produce floating, blur, or ghosting artifacts.OceanXL lacks explicit dynamic-object masks; motion segmentation or masking is left for future work.Source: §6.4 · p. 8

6.5 Practical Advantages for Real-Time Graphics and VR Applications

• Block-wise representation supports streaming and level-of-detail strategies, and compact per-chunk footprints are typically a few hundred megabytes even for large marine scenes.The discussion also notes that density control needs no auxiliary network during inference, easing integration with existing splatting rasterizers.Source: §6.5 · p. 8

6.6 Generalization Beyond Underwater Scenes

• The underwater-specific depth-banded densification component may generalize to other visibility-degraded settings, including fog, smoke, and atmospheric VFX backgrounds.Source: §6.6 · p. 8

7 Conclusion

• OceanXL combines BSP, underwater-aware lightweight density control, and selective Difference-of-Gaussians activation for efficient, compact reconstruction with high rendering fidelity.Across multiple underwater datasets, it improves efficiency-quality trade-offs over existing large-scale Gaussian splatting approaches.Source: §7 · p. 8
• Fixed depth stratification and the lack of explicit dynamic modeling are current limitations.Future work includes uncertainty-aware merging, physics-informed partitioning, and extension to temporal reconstruction.Source: §7 · p. 8