Sonar-neus:voxel-based efficient neural implicit surface reconstruction for forward-looking sonar.

Area of Science:

• Robotics and Autonomous Systems
• Oceanography and Marine Technology
• Computer Vision and Graphics

Background:

• Dense 3D reconstruction with forward-looking sonar (FLS) is crucial for ocean exploration.
• Current neural radiance field methods are too slow for real-world FLS 3D reconstruction.
• Existing methods suffer from slow training (hours) and inefficient sampling strategies.

Purpose of the Study:

• To develop a significantly faster and more efficient dense 3D reconstruction method for FLS data.
• To overcome the limitations of slow training and inefficient sampling in current neural radiance field approaches.
• To improve the quality and speed of 3D reconstruction for underwater environments.

Main Methods:

• Proposed a voxel-based neural implicit surface reconstruction approach for FLS.
• Replaced MLPs with geometry (SDF voxel grid) and appearance (feature voxel grid) representations.
• Introduced hierarchical sampling along the elevation arc and SDF Gaussian convolution for noise reduction.

Main Results:

• Achieved dense 3D reconstruction quality comparable to state-of-the-art methods.
• Reduced training time from 4 hours to just 10 minutes.
• Demonstrated superior performance over existing unsupervised dense FLS reconstruction techniques.

Conclusions:

• The proposed voxel-based method offers a substantial speed improvement for FLS 3D reconstruction.
• This advancement significantly enhances the feasibility of dense 3D reconstruction for real-world ocean exploration.
• The approach provides a more efficient and higher-quality solution for underwater mapping.