AquaFlow: A Monocular Gaussian Splatting SLAM for Underwater Streaming Reconstruction

TL;DR

AquaFlow achieves efficient underwater streaming reconstruction with monocular Gaussian Splatting SLAM, reducing localization error by 13.2% and improving PSNR by 4.74 dB.

cs.CV 🔴 Advanced 2026-08-24 30 views
Yingxiang Xu Kerui Ren Wenqi Guo Changjian Jiang Tao Lu Linning Xu Mulin Yu
underwater reconstruction streaming reconstruction 3D Gaussian Splatting SLAM deep learning

Key Findings

Methodology

AquaFlow employs a monocular Gaussian Splatting streaming reconstruction framework, integrating physics-inspired hybrid scene modeling. By fine-tuning a 3D vision foundation model on large-scale underwater data, it achieves robust pose tracking and pointmap estimation. It introduces a medium-guided incremental Gaussian initialization strategy, combining distance-aware neural Gaussians with a physical optical model to compensate for underwater image formation effects.

Key Results

  • AquaFlow was evaluated on 62 diverse underwater trajectory datasets, reducing localization error by 13.2% and improving PSNR by 4.74 dB compared to WaterSplat-SLAM.
  • It achieved high-precision scene reconstruction and rendering across various scales and complexities of underwater scenarios.
  • Ablation studies confirmed the effectiveness of combining physical models with neural Gaussians, significantly enhancing reconstruction quality.

Significance

This study is significant in the field of underwater 3D reconstruction, addressing visual degradation caused by light attenuation and scattering. AquaFlow not only improves reconstruction accuracy but also achieves breakthroughs in efficiency, providing new technical means for marine resource assessment and autonomous underwater navigation.

Technical Contribution

AquaFlow technically advances existing underwater reconstruction methods by combining physical models with neural networks, achieving efficient streaming reconstruction for the first time. Its innovative medium-guided incremental Gaussian initialization strategy and hybrid scene representation method offer new possibilities for precise modeling of underwater environments.

Novelty

AquaFlow is the first to apply monocular Gaussian Splatting SLAM to underwater scenarios, significantly enhancing reconstruction accuracy and efficiency by integrating physics-inspired optical models. It has a notable advantage in handling complex optical degradation compared to existing methods.

Limitations

  • In extreme lighting conditions, AquaFlow's reconstruction accuracy may decrease due to limited compensation for optical degradation.
  • For very complex underwater terrains, higher computational resources may be required for real-time reconstruction.

Future Work

Future work could explore more efficient computational methods to support real-time reconstruction of more complex scenes. Additionally, integrating multi-sensor data for multimodal fusion is a promising research direction.

AI Executive Summary

AquaFlow is a monocular Gaussian Splatting SLAM streaming reconstruction framework designed for underwater environments, addressing visual degradation issues caused by light attenuation and scattering. Existing methods perform poorly in underwater scenarios mainly due to a lack of modeling for complex optical effects. AquaFlow achieves efficient underwater reconstruction by fine-tuning a 3D vision foundation model and integrating physics-inspired hybrid scene modeling.

The method has been validated on large-scale underwater datasets, showing outstanding performance with a 13.2% reduction in localization error and a 4.74 dB improvement in PSNR. Its innovative medium-guided incremental Gaussian initialization strategy and hybrid scene representation method significantly enhance reconstruction accuracy and efficiency.

Despite significant progress in underwater reconstruction, AquaFlow still faces challenges in extreme lighting conditions. Future research could further enhance its performance and application scope through multimodal data fusion and more efficient computational methods.

Deep Analysis

Background

Underwater 3D reconstruction is crucial for marine resource assessment, infrastructure inspection, and autonomous navigation. However, existing methods perform poorly in underwater environments due to visual degradation caused by light attenuation and scattering. Recently, 3D Gaussian Splatting has made significant progress in terrestrial environments, but its application in underwater scenarios remains limited.

Core Problem

Visual degradation in underwater environments severely affects the accuracy of camera pose tracking and scene geometry reconstruction. Traditional methods often rely on offline processing, which cannot accommodate continuous video stream data. Achieving efficient and accurate streaming reconstruction in underwater environments is a pressing issue.

Innovation

AquaFlow combines physics-inspired optical models with neural Gaussian representations to achieve efficient streaming reconstruction in underwater environments for the first time. Its innovative medium-guided incremental Gaussian initialization strategy dynamically adjusts Gaussian parameters under incomplete observations, enhancing reconstruction accuracy.

Methodology

  • �� Fine-tune a 3D vision foundation model to adapt to underwater environments, enhancing the robustness of pose tracking and pointmap estimation.
  • �� Introduce a medium-guided incremental Gaussian initialization strategy, combining physical models to compensate for optical degradation.
  • �� Employ a hybrid scene representation method, integrating distance-aware neural Gaussians with a physical optical model for high-precision reconstruction.

Experiments

Evaluated on 62 diverse underwater trajectory datasets, including public benchmarks and web videos. Results show AquaFlow outperforms existing methods in localization accuracy and rendering quality. Ablation studies confirm the effectiveness of combining physical models with neural Gaussians.

Results

AquaFlow achieved significant improvements in localization error and PSNR, with a 13.2% reduction in localization error and a 4.74 dB improvement in PSNR. It performed excellently across various scales and complexities of underwater scenarios, demonstrating its potential in practical applications.

Applications

AquaFlow can be applied in marine resource assessment, infrastructure inspection, and autonomous underwater navigation. Its efficient reconstruction capability makes it highly promising for real-time applications.

Limitations & Outlook

In extreme lighting conditions, AquaFlow's reconstruction accuracy may decrease. Future research could enhance its performance and application scope through multimodal data fusion and more efficient computational methods.

Plain Language Accessible to non-experts

Imagine you're filming underwater, but the video quality is poor due to water refraction and light attenuation. AquaFlow acts like a super filter that can fix these videos in real-time. It analyzes water properties and adjusts colors and details in the video, like giving it a makeover. This way, you can see a clear underwater world, just like in an aquarium.

ELI14 Explained like you're 14

Hey there! Imagine you're filming underwater, but the video looks blurry because water makes light act weird. AquaFlow is like a super-smart filter that can fix these videos in real-time. It recognizes water properties and adjusts colors and details, letting you see a clear underwater world, just like in an aquarium! Cool, right?

Glossary

Gaussian Splatting

A technique for 3D scene reconstruction using Gaussian distributions to represent points in the scene.

Used to represent the geometric structure of underwater scenes.

SLAM

Simultaneous Localization and Mapping, a technique for estimating camera poses and building environment maps simultaneously.

Used in AquaFlow for pose tracking in underwater environments.

PSNR

Peak Signal-to-Noise Ratio, a metric for measuring image or video quality; higher values indicate better quality.

Used to evaluate AquaFlow's rendering quality.

Optical Model

A physical model used to simulate light propagation in a medium.

Used to compensate for optical degradation in underwater images.

Neural Gaussian

A Gaussian representation combined with neural networks for dynamically adjusting scene parameters.

Used to improve reconstruction accuracy and efficiency.

Open Questions Unanswered questions from this research

  • 1 How to improve reconstruction accuracy in extreme lighting conditions? Current models have limited compensation for optical degradation.
  • 2 How can multimodal data fusion enhance the robustness of underwater reconstruction?
  • 3 How to reduce computational resource requirements in complex underwater terrains?

Applications

Immediate Applications

Marine Resource Assessment

AquaFlow can be used for assessing marine resources by providing detailed seabed topography through high-precision underwater reconstruction.

Long-term Vision

Autonomous Underwater Navigation

With real-time efficient underwater reconstruction, AquaFlow can be used for autonomous underwater navigation, enhancing the navigation capabilities of underwater robots in complex environments.

Abstract

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.

cs.CV