Abstract
Underwater 3D reconstruction is challenged by light scattering and absorption, which lead to significant color casts and low contrast. This paper presents a unified pipeline that integrates advanced image enhancement, intelligent keyframe selection, and Instant-NGP to achieve high-quality 3D models. The proposed system significantly reduces reconstruction time from hours to approximately one minute while maintaining competitive visual fidelity compared to state-of-the-art physics-based models.
Introduction
Reconstructing 3D scenes from underwater imagery is notoriously difficult due to the properties of water as a scattering medium. Selective light absorption causes significant color shifts, while scattering results in blur and reduced contrast. While Neural Radiance Fields (NeRF) have revolutionized 3D reconstruction in clear air, their application underwater is hindered by poor raw image quality and high computational costs. Existing solutions like Seathru-NeRF, while effective, require extensive training times that limit practical deployment.
Proposed Methodology
The authors propose a multi-stage system designed for both speed and quality in scattering environments.
Underwater Image Enhancement
To provide a stable foundation for reconstruction, the system first processes video frames using a hybrid enhancement approach. It combines Contrast Limited Adaptive Histogram Equalization (CLAHE) for local contrast improvement with a Bayesian Retinex algorithm. This decomposition into reflectance and illumination components effectively removes color casts and restores natural scene appearance, ensuring spatial and temporal consistency for the subsequent 3D modeling phase.
Keyframe Filtering and Reconstruction
To optimize computational efficiency and avoid artifacts from dynamic objects, a weighted importance parameter ($I$) is used to select representative keyframes. This parameter accounts for both camera pose (angular difference and displacement) and image sharpness (calculated via a Laplacian operator), ensuring that blurred frames from turbulence or camera shake are excluded.
The final reconstruction utilizes Instant-NGP (Instant Neural Graphics Primitives). By employing multi-resolution hash encoding, the system allows a smaller neural network to represent the scene, drastically reducing training and rendering times compared to traditional NeRF architectures.
Experimental Results
The system was evaluated using video data from underwater diving expeditions captured with GoPro HERO BLACK cameras at $1280 \times 720$ resolution. Two distinct scenarios were tested, consisting of 241 and 231 images respectively.
Performance and Efficiency
The proposed system achieved a massive leap in efficiency:
- Speed: Reconstruction was completed in approximately 1 minute, a 420x speedup over Seathru-NeRF (7 to 20 hours).
- Fidelity: The method achieved a PSNR of 18.43, outperforming Seathru-NeRF's score of 16.54.
- Structural Similarity: While Seathru-NeRF maintained a slight edge in SSIM (0.6958 vs. 0.6610) due to its explicit light transport modeling, the proposed method offers a far superior trade-off for rapid applications.
Conclusion
This work demonstrates that an "enhancement-first" approach combined with efficient neural primitives can overcome the traditional bottlenecks of underwater 3D reconstruction. Future work will focus on integrating underwater light physics directly into the neural volume rendering process to further reduce artifacts and improve color fidelity without sacrificing the current speed advantages.
Summary updated