Aqua-Splat: Physically-Informed Sonar-Camera Gaussian Splatting for Underwater 3D Reconstruction
Zijie Ling, Yunxuan Feng, Ao Meng, Renxiang Xiao, Shu Pan, Wenjie Lu, Liang Hu
AI summary
Problem
Standard Gaussian Splatting fails underwater due to light attenuation and scattering, while existing sonar-camera fusion methods either ignore acoustic wave physics or lack real-time rendering speeds.
Approach
Aqua-Splat transforms 3D Gaussians into a polar coordinate system and applies a physically-informed volume rendering pipeline that models sonar wave propagation, guided by a sonar-specific densification strategy for rapid scene optimization.
Key results
- Physically-informed polar-domain sonar rendering model
- GPU-accelerated volume rendering achieving over 120 FPS
- Sonar-guided densification strategy for faster convergence
- Superior geometric and photometric reconstruction over camera-only and prior multimodal baselines
Why it matters
Provides underwater robots and marine researchers with a fast, robust perception tool for high-fidelity mapping in optically degraded environments.
Abstract
Differentiable Gaussian Splatting (GS) has emerged as a powerful paradigm for scene representation, enabling efficient rendering and real-time editing. However, existing GS- based methods, which rely mainly on clear visual images, perform poorly in underwater environments due to camera distortions such as light absorption and backscattering. In contrast, acoustic sensors like Forward Looking Sonar (FLS) offer superior pene- tration and robustness in such conditions. To leverage the com- plementary merits of visual and FLS images, we propose a novel GS framework customized for underwater scenarios, termed Aqua-Splat, for robust and accurate underwater perception. It ensures physically consistent reconstruction by incorporating the sonar wave propagation modeling in the image formation process. Moreover, we propose a volume rendering technique for sonar image synthesis, achieving similar speed to visual ren- dering. Additionally, we introduce a sonar-guided densification strategy to optimize the scene representation. Through extensive experiments on both simulated and datasets from the lab pool, we demonstrate that Aqua-Splat significantly improves image synthesis and 3D scene reconstruction in challenging underwater environments, outperforming existing methods in terms of both geometric accuracy and photometric fidelity. The code of Aqua- Splat will be open-sourced later for the community.