Related Experiment Video
Updated: Jun 12, 2026
13:01
Using Light Sheet Fluorescence Microscopy to Image Zebrafish Eye Development
Published on: April 10, 2016
Subaquatic neural view synthesis with depth-guided refinement and multi-scale information fusion
Huilin Ge1, Zheng Wang1, Bingying Hu1
1Jiangsu University of Science and Technology, No. 666 Changhui Road, Dantu New District, Zhenjiang, Jiangsu, 212100, China.
Summary
This study introduces a novel neural rendering framework that improves underwater 3D reconstruction and novel-view synthesis by combining physics-based models with depth-aware optimization, enhancing realism and robustness in challenging aquatic environments.
Area of Science:
- Computer Vision
- Computer Graphics
- Optical Engineering
Background:
- Underwater images suffer from significant degradation due to light attenuation and backscattering.
- This degradation severely impacts 3D reconstruction and the synthesis of novel views from existing images.
Purpose of the Study:
- To develop a unified neural rendering framework for high-fidelity underwater novel-view synthesis.
- To address the challenges of light attenuation and backscattering in underwater imaging.
Main Methods:
- Coupling a signed distance function (SDF) with a physics-based underwater image-formation (UIF) model.
- Implementing a depth-aware optimization scheme for improved geometric accuracy and color consistency.
- Utilizing multi-resolution hash encoding, numerical gradients, and four pseudo-depth regularizers for enhanced structural stability and surface-normal continuity.
Main Results:
- Demonstrated superior performance over existing methods on SeaThru-NeRF and S-UW datasets.
- Achieved significant improvements in novel-view synthesis, especially at long ranges and under strong scattering conditions.
- Showcased enhanced robustness and realism in generated underwater scenes.
Conclusions:
- The proposed physics-guided neural rendering framework effectively overcomes underwater image degradation.
- The integration of physics-based priors is crucial for achieving realistic and robust neural rendering in challenging underwater environments.