AquaFlow appeared on arXiv on August 24, 2026 as arXiv:2608.22906, jointly authored by Zhejiang University, Shanghai AI Laboratory, Shanghai Jiao Tong University, Tsinghua University, University of Hong Kong, and the Chinese University of Hong Kong. The problem it addresses is a single sentence: real-time 3D scene reconstruction from monocular video underwater.
Headline results versus WaterSplat-SLAM, its direct comparison baseline, across 62 underwater trajectories:
- Average localization error reduced by 13.2%
- PSNR improved by 4.74 dB
- Why underwater SLAM is an order of magnitude harder than land SLAM: ambient-light attenuation, time-varying turbidity, backscatter, and robot motion from currents each break a core assumption of terrestrial 3DGS pipelines.
- A three-part pipeline: (1) a 3D vision foundation model fine-tuned on large-scale underwater data, (2) media-guided incremental initialization, (3) a hybrid scene representation combining structured distance-conditioned neural Gaussians with a physics-inspired optical model.
- Results: 13.2% lower absolute trajectory error (ATE) and 4.74 dB higher PSNR than WaterSplat-SLAM over 62 underwater trajectories (~35 public benchmark + ~27 in-the-wild web videos).
- Limitation: the optical model is assumed known/estimable; sudden turbidity changes or out-of-distribution conditions can trigger pose drift. The authors explicitly flag out-of-distribution robustness as unverified.
- Public underwater benchmark datasets;
- In-the-wild underwater videos scraped from the web (YouTube, TikTok, ocean documentaries);
- Total: 62 underwater trajectories at different scales.
- Public benchmarks (~35): from public underwater SLAM datasets;
- In-the-wild web videos (~27): unknown camera parameters, unknown turbidity, unknown lighting, unknown depth.
- arXiv:2608.22906 (online 2026-08-24): AquaFlow: A Monocular Gaussian Splatting SLAM for Underwater Streaming Reconstruction
- Authors: Yingxiang Xu, Kerui Ren, Wenqi Guo, Changjian Jiang, Tao Lu, Linning Xu, Mulin Yu
- Institutions: Zhejiang University + Shanghai AI Laboratory + Shanghai Jiao Tong University + Tsinghua University + University of Hong Kong + CUHK
- arXivDaily industry trends repost, 2026-08-25
- Same-period coverage: LiteGS (M Threads, ECCV 2026), 4DAnyone (SIGGRAPH Asia 2026), tiled Gaussian reconstruction for large-scale scenes
A 4.74 dB gain in image reconstruction is a visually obvious tier of improvement; 13.2% lower localization error is a publishable SOTA margin in SLAM. Achieving both simultaneously, validated repeatedly on a public dataset, is the kind of result seen at ICCV/ECCV-level review.
But AquaFlow's key insight is not the numbers—it rewrites the 3DGS SLAM pipeline around a simple physical assumption: underwater light is medium-dependent.
Key points
1. Water Is Not Air: Why Land Assumptions Break
Moving a 3DGS SLAM system from land to water breaks the implicit assumptions of "appearance consistency across adjacent frames + stable camera pose":
| Assumption | Land | Underwater | |---|---|---| | Adjacent frames look consistent | Yes | No (light attenuates with distance) | | Surface color is stable | Yes | No (turbidity varies over time) | | Backscatter is negligible | Yes | No (suspended particles bounce light back) | | Camera pose is stable | Yes | Partially (currents push the robot) | | Textures support geometry | Yes | No (color attenuation masks structure) |
Each failure alone is manageable; combined, every stage of the traditional pipeline (point cloud initialization → Gaussian fitting → incremental update) drifts.
AquaFlow's engineering philosophy: stop forcing land assumptions; explicitly model the underwater physical process in the pipeline.
2. Component One: Fine-Tuned 3D Vision Foundation Model
The pipeline's first building block is a 3D vision foundation model fine-tuned on large-scale underwater data (the paper does not name the model; structure suggests a DUSt3R/MASt3R-style pointmap regression model). The goal: make its camera pose + pointmap outputs robust to underwater visual degradation. Land-trained models misinterpret water's color shifts as object motion, producing pseudo pose drift.
Fine-tuning data composition:
3. Component Two: Media-Guided Incremental Initialization
Land 3DGS SLAM initializes point clouds via depth-map back-projection + multi-view fusion. Underwater, depth estimates get corrupted by refraction and fusion by distance-dependent color attenuation.
AquaFlow's two steps:
1. Media-guided filtering: only regions consistent with the current frame's optical model prediction enter the point cloud; 2. Incremental updating: reliable regions from new frames are added progressively rather than reconstructed all at once.
The essence: only build geometry where optics are consistent; leave the rest to the hybrid representation to compensate. It is a "media-aware switch" in the SLAM pipeline.
4. Component Three: Hybrid Scene Representation
Traditional 3DGS encodes each Gaussian with position, covariance, opacity, and spherical harmonics for color. Underwater, two problems arise: color is polluted by distance-dependent attenuation, and the SH coefficients are forced to encode water's medium effects instead of pure reflectance.
AquaFlow splits the representation into two parallel parts:
| Component | Role | Encodes | |---|---|---| | Structured distance-conditioned neural Gaussians | Geometry | Surface position, covariance, opacity | | Physics-inspired optical model | Imaging | Attenuation and scattering across propagation distances |
Engineering consequence: color variation no longer has to be crammed into geometry parameters; pose and appearance can be constrained separately during streaming updates. The murkier the water, the bigger the payoff—traditional 3DGS in turbid water inevitably encodes water's attenuation as reflectance change, producing bloated textures and accumulating pose error.
5. The 62 Trajectories
Experiments cover 62 underwater trajectories:
The web portion represents real deployment complexity. Outperforming WaterSplat-SLAM on both standardized and real-world data is what makes the result convincing.
6. How to Read the Numbers
| Time | Work | Dataset | Improvement | |---|---|---|---| | 2024 | WaterSplat-SLAM | Public underwater benchmarks | Land 3DGS paradigm first moved underwater | | 2026 Q3 | AquaFlow | 62 underwater trajectories | 13.2% ATE reduction + 4.74 dB PSNR |
The team is candid:
> These results cover public benchmarks and web-collected videos, but are averages over the paper's specified data and baselines. Once turbidity, lighting, depth, or camera motion exceed the training distribution, the same gains cannot be assumed.
In other words: SOTA on the reported 62 trajectories, but cross-scene extrapolation carries risk.
7. Why This Matters for Marine Robotics
Three concrete use cases benefit from monocular, real-time, centimeter-scale underwater reconstruction:
1. Aquaculture (marine ranch) inspection: periodic ROV patrols of net cages, fish monitoring, and net-damage detection; sonar-plus-camera solutions are costly and inefficient. 2. Seabed terrain mapping: high-resolution bathymetry is strategic infrastructure; a 3DGS SLAM paradigm lets monocular-camera ROVs perform centimeter-level reconstruction. 3. Deep-sea aquaculture + subsea cable patrol: low-cost monocular video reconstruction plus automated anomaly detection is the core need for both.
8. Limitations: The Known-Optical-Model Assumption
AquaFlow assumes the optical model (turbidity, attenuation, scattering parameters) is known or estimable in advance. If turbidity suddenly doubles—rain, a current shift—media-guided initialization may misread "suddenly murky" as "the surface changed," triggering pose drift. Future steps likely include online optical parameter estimation, dynamic-scene extension (moving organisms, current disturbance), and fusion with sonar/IMU to fall back on acoustic SLAM in extreme turbidity.
9. Context in the 3DGS Timeline
| Time | Work | Contribution | |---|---|---| | 2023 Q4 | 3DGS (INRIA) | Replaced NeRF with Gaussian ellipsoids | | 2024 H1 | Real-time 3DGS rendering (NVIDIA) | Training from hours to minutes | | 2024 Q4 | 3DGS SLAM (multiple works) | 3DGS embedded in SLAM pipelines | | 2025 | Large-scale tiled 3DGS | City/block-scale reconstruction | | 2026 Q1 | 4DAnyone (Ant + Zhejiang) | Monocular video → 4D dynamic human reconstruction | | 2026 Q3 | AquaFlow | Monocular video → underwater 3DGS SLAM |
AquaFlow occupies the hardest quadrant—outdoor/extreme, partially dynamic—and opens 3DGS to oceans, endoscopic imaging, and extreme environments (space, polar).
10. Takeaways for 3DGS Startups
1. Scene-specific physics adaptation is the moat. The base paradigm (Gaussian ellipsoids + SH) is mature; differentiation comes from adapting physical assumptions per environment—exactly what media guidance + the optical model do for water. 2. Multi-modal fusion is next. Camera + IMU + sonar + depth fusion moves 3DGS from visual reconstruction to multi-modal reconstruction—critical for underwater, indoor, glare, and fog. 3. Open datasets are key to industrialization. The 62-trajectory dataset, if released, would be a low-cost iteration starting point; the institutional lineup (Zhejiang, Shanghai AI Lab, CUHK) makes open-sourcing plausible, though not confirmed.
11. Conclusion
AquaFlow answers a question bigger than underwater reconstruction: when a paradigm (3DGS) meets a physical environment change (air → water), do you rewrite the paradigm, or rewrite its physical assumptions? Past three years' mainstream answer was to rewrite the paradigm—NeRF → 3DGS, dense → tiled, single-frame → dynamic. AquaFlow offers a lighter path: keep 3DGS intact, but explicitly model the medium's physics. That path naturally extends to deep sea, endoscopy, space, and polar environments.
Three numbers to watch over the next 12 months: whether the 62-trajectory dataset is open-sourced; whether a Chinese marine robotics company integrates AquaFlow into ROV control stacks; and whether non-underwater variants (endoscopic, foggy driving, space) adopt the physics-medium-guided paradigm. Any one happening would turn 3DGS from a visual reconstruction tool into infrastructure for extreme-environment reconstruction.
---
Sources