AquaGS: Fast Underwater Scene Reconstruction with SfM-Free Gaussian Splatting

TL;DR

AquaGS tackles the challenge of underwater 3D reconstruction without Structure-from-Motion by combining a fast learning-based MVS initialization with Gaussian Splatting for surface geometry and a NeRF-based model for the translucent underwater medium. The method explicitly separates object and medium contributions via SeaThru-inspired rendering, and jointly optimizes camera poses and scene attributes using a reconstruction and transparency loss. Key contributions include stable Gaussian initialization through a pre-trained network with confidence-based downsampling, medium removal via separate object/medium parameters, and real-time rendering capabilities demonstrated on underwater robots, achieving high-precision reconstructions in seconds from as few as three images. This approach significantly enhances practical underwater perception for robotic platforms, enabling rapid scene understanding where traditional SfM-based methods struggle or fail, and points toward efficient embedded deployment through memory and computation optimizations.

Abstract

Underwater scene reconstruction is a critical tech-nology for underwater operations, enabling the generation of 3D models from images captured by underwater platforms. However, the quality of underwater images is often degraded due to medium interference, which limits the effectiveness of Structure-from-Motion (SfM) pose estimation, leading to subsequent reconstruction failures. Additionally, SfM methods typically operate at slower speeds, further hindering their applicability in real-time scenarios. In this paper, we introduce AquaGS, an SfM-free underwater scene reconstruction model based on the SeaThru algorithm, which facilitates rapid and accurate separation of scene details and medium features. Our approach initializes Gaussians by integrating state-of-the-art multi-view stereo (MVS) technology, employs implicit Neural Radiance Fields (NeRF) for rendering translucent media and utilizes the latest explicit 3D Gaussian Splatting (3DGS) technique to render object surfaces, which effectively addresses the limitations of traditional methods and accurately simulates underwater optical phenomena. Experimental results on the data set and the robot platform show that our model can complete high-precision reconstruction in 30 seconds with only 3 image inputs, significantly enhancing the practical application of the algorithm in robotic platforms.

Figures (6)

• Figure 1: Overview of AquaGS. In the underwater environment, our vehicle captures real-time images, sequentially performing Fast Initialization, SeaThru Rendering and Joint Optimization, producing an accurate restoration of underwater objects.
• Figure 2: Testing Bed Settings. We integrate our model into robot. The reconstruction results are achieved on the computer.
• Figure 3: Visual Comparisons of Underwater Scene Rendering between AquaGS and Various Baseline Methods. AquaGS achieves excellent results, while 3DGS and InstantSplat cannot effectively deal with the effects of underwater media on objects.
• Figure 4: Scene Separation, on ‘Panama’. In the bottom left corner, the depth map and medium image are shown, while the bottom right corner displays a magnified view of the details.
• Figure 5: Robot Experiment Results on ‘Black Barricade’. Upon visual examination, our method for separating the medium and objects matches the performance of SeaThru-NeRF, with both approaches delivering exceptional outcomes.