Source-linked AI summary
AquaFlow: A Monocular Gaussian Splatting SLAM for Underwater Streaming Reconstruction
Yingxiang Xu, Kerui Ren, Wenqi Guo, Changjian Jiang, Tao Lu, Linning Xu, Mulin Yu
TL;DR
Underwater attenuation, scattering, and degraded visual correspondences challenge monocular streaming reconstruction. AquaFlow adapts a 3D vision foundation model and combines medium-guided Gaussian initialization with a hybrid optical-neural representation; across 62 trajectories, it achieves state-of-the-art tracking and rendering performance.
Problem
Underwater optical degradation degrades camera pose tracking and distorts scene geometry, limiting monocular 3DGS streaming reconstruction.
Method
AquaFlow fine-tunes an underwater 3D vision foundation model and combines medium-guided Gaussian initialization with distance-aware neural Gaussians and explicit optical modeling.
Results
AquaFlow achieves the lowest overall mean trajectory localization error of 0.401 m and state-of-the-art reconstruction performance across evaluated underwater datasets.
Takeaways & Limitations
The results support AquaFlow as a scalable monocular framework for high-fidelity underwater streaming reconstruction across diverse real-world trajectories.
Abstract
from arXiv · showhide
Recent monocular 3D Gaussian Splatting (3DGS) streaming reconstruction methods have achieved impressive performance by balancing reconstruction quality and efficiency. However, extending these frameworks to underwater scenes remains challenging due to severe visual degradation, such as light attenuation and scattering, which degrades camera pose tracking and distorts scene geometry. To address these challenges, we propose AquaFlow, a monocular Gaussian Splatting streaming reconstruction framework for efficient and high-fidelity underwater reconstruction. Specifically, AquaFlow fine-tunes a 3D vision foundation model on large-scale underwater data for robust pose and pointmap estimation, and introduces a medium-guided incremental Gaussian initialization strategy for streaming mapping. Furthermore, we develop a streaming-compatible hybrid scene representation that integrates structured, distance-conditioned neural Gaussians with a physics-inspired optical model to compensate for underwater image formation effects, enabling accurate scene reconstruction. We evaluate AquaFlow on a comprehensive dataset of 62 diverse underwater trajectories, collected from both public benchmarks and in-the-wild web videos across various scales. Extensive experiments demonstrate that AquaFlow achieves state-of-the-art tracking and rendering performance, reducing average localization error by 13.2% and improving PSNR by 4.74 dB compared to WaterSplat-SLAM.
I. INTRODUCTION
AquaFlow addresses the difficulty of scalable monocular underwater reconstruction by combining underwater-adapted geometric priors with streaming Gaussian mapping and physics-guided scene modeling.
- Underwater attenuation, scattering, and weak textures make stable camera-pose estimation and accurate continuous-video reconstruction difficult.
- Existing underwater 3DGS methods are mainly offline, requiring complete sequences, pre-estimated poses, sparse point clouds, and costly optimization.
- AquaFlow fine-tunes a 3D vision foundation model on subsea data to provide robust geometric priors for pose tracking and dense point-map estimation.
- Its streaming-compatible hybrid representation combines distance-aware neural Gaussians with an explicit underwater optical model to separate scene appearance from attenuation and backscattering.
- The framework is evaluated on 62 diverse real-world underwater sequences and reports state-of-the-art camera tracking and high-fidelity reconstruction.
B. Offline Underwater 3D Gaussian Splatting Reconstruction
AquaFlow targets limitations of offline and general streaming reconstruction by combining monocular streaming estimation with underwater-aware medium modeling and geometric priors.
- Offline underwater 3DGS methods optimize physical light transport and Gaussian primitives from pre-computed SfM inputs to produce photorealistic novel views.
- General streaming reconstruction methods face underwater pose drift, erroneous Gaussian growth, and suboptimal scene representations.
- AquaFlow jointly estimates camera trajectories and underwater attenuation and backscatter parameters from a single monocular RGB video.
- Its pipeline combines an underwater-adapted 3D geometric prior, tracking and global optimization, hybrid scene-medium representation, and joint optimization.
- The adapted foundation model is trained using a diverse underwater 3D dataset spanning synthetic and real environments with geometric supervision.
2) Model finetuning:
AquaFlow adapts MASt3R for underwater geometry by fine-tuning it on calibrated underwater image pairs with metric-scale supervision and feature-matching objectives.
- MASt3R is selected for rapid inference, stable small-baseline geometry estimation, and compatibility with SLAM front-end tracking.
- Given two uncalibrated underwater images, the model predicts 3D geometry, per-pixel descriptors, and confidence maps in one forward pass.
- AquaFlow fine-tunes MASt3R on 224K calibrated underwater two-view image pairs for metric-scale 3D predictions.
- The training objective combines metric pointmap regression with log-confidence regularization and InfoNCE feature matching over 3D-near correspondences.
B. Tracking and GLobal Optimization
The tracking frontend uses foundation-model pointmaps, descriptors, and confidence estimates to establish correspondences and optimize relative and global camera poses online.
- For each incoming frame, AquaFlow compares it with the latest keyframe to obtain dense pointmaps, descriptors, and confidence maps.
- Relative transformations are estimated by minimizing a confidence-weighted pointmap alignment residual with a robust Huber loss.
- Keyframes are inserted when reliable match counts fall below a tracking threshold, triggering joint optimization of active keyframe poses.
- Mapping frames are selected when high-confidence matched-point parallax exceeds a baseline threshold, initiating new Gaussian insertion.
- Common frames that meet neither criterion only update and refine existing Gaussian parameters.
C. Hybrid Scene Representation
AquaFlow represents underwater scenes with structured Gaussian groups and observation-conditioned neural decoding. Distance and viewing direction guide residual updates to Gaussian geometry and appearance, improving modeling of range-dependent underwater effects.
- Each Gaussian explicitly learns position and opacity, while rotation, scale, and zeroth-order color are decoded as residual updates.The decoder operates on structured Gaussian tokens and compensates for coarse streaming initialization.
- Observation distance and viewing direction are concatenated into a condition for distance-aware Gaussian parameter prediction.The distance approximates light propagation through water, while the viewing direction captures observation geometry.
- Sinusoidal encoding of the observation condition captures complex, range-dependent underwater imaging phenomena that are difficult to model explicitly.The encoding uses L frequency bands.
- Voxelized Gaussian groups share structural context and are processed by LGPD with validity masking and local context aggregation.Each group contains at most K primitives, and the network aggregates neighboring Gaussian information.
- Residual corrections update Gaussian rotation, scale, and color, improving the stability of streaming optimization.
2) Underwater Medium Modeling:
AquaFlow models underwater image formation by combining attenuated scene radiance with accumulated backscatter. Its medium parameters can depend on camera orientation, enabling directional modeling of underwater appearance.
- The observed underwater color combines attenuated scene radiance with accumulated backscatter.The scene radiance is rendered from the Gaussian map, while depth along the camera ray conditions the medium effects.
- AquaFlow predicts attenuation coefficients from camera orientation rather than assuming isotropic medium parameters.The camera rotation is flattened into a pose vector and sinusoidally encoded before coefficient decoding.
- The backscatter formulation uses learnable channel-wise vectors for asymptotic intensity, accumulation rate, near-distance residual, and residual decay.
3) Medium-guided Incremental Gaussian Initialization:
AquaFlow initializes new Gaussians from medium-reduced observations rather than degraded underwater images. A structure-aware insertion probability identifies regions requiring new primitives, while residual color updates remain learnable during optimization.
- New Gaussians are initialized using medium-reduced observations produced by the online learning medium model.This avoids directly using degraded underwater images for newly observed scene regions.
- The medium-reduced image is recovered using estimated attenuation, backscatter, and depth rendered from the current Gaussian map.The recovered image provides cleaner appearance cues for Gaussian initialization.
- A structure-aware insertion probability compares the medium-reduced observation with the current medium-free rendering.The probability is computed with a Laplacian-of-Gaussian operator.
- Pixels with Pa(u, v) > τa receive new Gaussians whose base colors come from the medium-reduced image.The base color remains fixed while neural enhancement predicts residual color updates during later optimization.
D. Loss Function
AquaFlow jointly optimizes Gaussian attributes, learned features, decoder parameters, and underwater medium parameters through differentiable rendering. Its objective combines color, depth, and scale regularization while masking outliers and weighting central image regions.
- AquaFlow jointly optimizes Gaussian attributes, local and global features, LGPD parameters, and medium parameters through differentiable rendering.
- A radial decay kernel weights spatial loss terms to compensate for boundary lens distortions and prioritize central-view fidelity.
- An adaptive binary validity mask filters common-frame outliers using weighted rendered-image residuals, while keyframes and mapping frames use an all-ones mask.
- The mask is applied to rendered images, ground-truth images, rendered inverse depth, and monocular depth priors from the underwater 3D vision foundation model.
- The total objective combines color, depth, and Gaussian scale regularization terms with balancing weights for structural similarity, depth consistency, and scale regularization.
IV. EXPERIMENT
AquaFlow is evaluated on diverse underwater sequences using reconstruction, tracking, and efficiency metrics, with qualitative comparisons covering distant scenes and complex seafloor structures.
- Test Datasets: The test suite combines public and in-the-wild underwater sequences spanning small- and large-scale scenes.The public subset includes SeaThru-NeRF, S-UW, Canyons, RedSea, and UW-Stereo-VI sequences.
- Evaluation Metrics: Reconstruction quality is measured by PSNR, SSIM, and LPIPS, while tracking uses aligned ATE RMSE.Efficiency is assessed through reconstruction time and per-frame processing latency.
- Quantitative Evaluation: Table II reports quantitative reconstruction results across six real-world underwater benchmarks.The table reports PSNR, SSIM, and LPIPS, with higher PSNR/SSIM and lower LPIPS indicating better performance.
- Qualitative Evaluation: AquaFlow’s qualitative comparisons cover distant underwater scenes, intricate seafloor details, and complex subsea topologies.Examples include ground markings, discarded vehicle textures, underwater caves, shipwreck ruins, and canyons.
2) Baselines:
AquaFlow is compared with state-of-the-art tracking, rendering, and efficiency baselines using a domain-adapted foundation model and hybrid underwater scene representation. It achieves state-of-the-art reconstruction on five datasets while preserving fine details and structural boundaries.
- Baselines: AquaFlow is benchmarked against optical-flow, foundation-model, underwater-specific, streaming, and offline reconstruction baselines.The comparisons cover camera tracking, rendering quality, runtime, and streaming efficiency.
- 3D Vision Foundation Model Training Data: AquaFlow fine-tunes MA3StR on underwater data combining synthetic geometry with real-world optical degradation.The training set contains 224,273 valid image pairs curated from five public underwater datasets.
- Efficiency Evaluation: AquaFlow reports quantitative efficiency using scene training time, training throughput, and Gaussian count.Table III compares these measures across methods, with lower time and Gaussian count and higher throughput preferred.
- Evaluation Protocol: AquaFlow evaluates localization with ATE RMSE and reconstruction with PSNR, SSIM, and LPIPS across underwater benchmarks.The evaluation includes held-out frames for pose and trajectory assessment.
- Reconstruction Results Analysis: AquaFlow achieves state-of-the-art reconstruction performance on five datasets, improving PSNR and SSIM while reducing LPIPS.It is slightly below offline UW-GS on SeaThru-NeRF, whose scenes contain around 20 images focused on localized coral environments.
- Reconstruction Results Analysis: AquaFlow preserves long-range objects, fine seafloor details, sharp boundaries, and intricate structures in qualitative comparisons.The method combines explicit underwater light-transport modeling, distance-conditioned Gaussians, and underwater geometric priors.
2) Tracking Results Analysis:
AquaFlow achieves strong underwater tracking and reconstruction while maintaining streaming efficiency. Ablations show that its geometric priors and medium-aware scene modeling are important for preserving pose accuracy, geometry, appearance, and detail.
- Tracking Results Analysis: 0.401 m mean localization error is the lowest overall among the compared SLAM methods.AquaFlow also achieves top tracking performance on four of six real-world datasets.
- Tracking Results Analysis: AquaFlow maintains comparable or superior runtime efficiency to streaming baselines while achieving significantly higher reconstruction quality.It also operates significantly faster than the offline UW-GS baseline, excluding UW-GS preprocessing time.
- Ablation Study: Removing any core component causes consistent performance drops, while the full model achieves optimal performance across all reported metrics.The ablations remove medium-aware initialization, distance embeddings, underwater fine-tuning, or LGPD.
- Ablation Study: Without medium-aware initialization, severe attenuation and low contrast produce sparse seeds and insufficient Gaussian modeling, causing structural scene loss.The qualitative comparison reports structural geometric errors or scene blurring when specific modules are disabled.
- Ablation Study: Removing underwater fine-tuning corrupts depth boundaries and collapses geometry, while omitting distance embeddings shifts colors and reduces synthesized-view contrast.Replacing LGPD with an MLP further weakens physical decoupling and blurs fine details and high-frequency textures.
- Tracking Results Analysis: Evaluations across 62 real-world underwater trajectories show state-of-the-art localization and reconstruction performance with high time efficiency.The trajectories span diverse underwater environments and support comparisons with offline and streaming baselines.