Capturing Motion,
Crafting Stories

A research team led by Professor Liang Hu at Harbin Institute of Technology, Shenzhen published a paper in IEEE Robotics and Automation Letters (IEEE RA-L) titled “Aqua-Splat: Physically-Informed Sonar-Camera Gaussian Splatting for Underwater 3D Reconstruction.”

The paper presents Aqua-Splat, a sonar–vision method for underwater reconstruction and novel-view synthesis based on a 3D Gaussian representation. Aqua-Splat incorporates the physical imaging principles of forward-looking sonar into 3D Gaussian Splatting, enabling sonar and visual data to jointly optimize the same underwater 3D scene and improve the physical consistency of the reconstruction. Experiments conducted in simulated environments and a laboratory pool demonstrated improvements in geometric reconstruction accuracy and photometric fidelity, while sonar image rendering exceeded 120 FPS.

NOKOV motion capture system, comprising underwater motion capture cameras, provided ground-truth poses of a BlueROV2 underwater robot equipped with a ZED2 stereo camera and an Oculus m750d forward-looking sonar. Combined with the sensor extrinsic parameters, these measurements enabled the researchers to determine the position and orientation at which each visual and sonar frame was captured, supporting the construction of an underwater vision–sonar–pose multimodal dataset.

The pose data established accurate spatial correspondence between the visual and sonar frames and provided the sensor-pose information required for coordinate transformations, sonar–vision joint optimization, underwater 3D reconstruction, and novel-view synthesis in Aqua-Splat.

Paper Information

Paper: Aqua-Splat: Physically-Informed Sonar-Camera Gaussian Splatting for Underwater 3D Reconstruction

Journal: IEEE Robotics and Automation Letters, 2025

DOI: 10.1109/LRA.2025.3615528

Original Paper: View the paper on IEEE Xplore