High-Fidelity 3D Seafloor Scene Reconstruction Based on Cross-dimensional Gaussian Normal Transition Field
-
摘要: 海洋科学勘测与水下环境探查等领域对海底高保真场景重建的需求日益提升, 三维高斯溅射技术作为一种先进的显式场景表示方法, 在海底环境重建与新视图合成任务中展现出显著应用前景。然而, 受水下成像介质模糊效应等影响, 重建结果常出现介质伪影与结构失真等缺陷, 严重制约其在实际水下复杂环境下的适用性。为解决上述问题, 文中提出一种基于跨维高斯法向跃迁场的海底三维高保真场景重建方法, 首先建立高斯基元的跨维映射与法向跃迁系统, 提升其对复杂结构的精细化几何建模能力; 其次提出高斯不透明度加权滤波模型, 抑制由介质模糊效应引发的重建伪影; 最后在多种水下场景的实验结果表明该方法具备高效处理复杂水下场景重建与新视图合成的能力。Abstract: The demand for high-fidelity seafloor scene reconstruction is growing in fields such as marine scientific surveying and underwater environmental exploration. As an advanced explicit scene representation method, three-dimensional(3D) Gaussian Splatting holds significant application potential for seafloor reconstruction and novel view synthesis. However, influenced by factors such as blurring effects caused by underwater imaging media, the results often exhibit defects including medium-induced artifacts and structural distortions, severely limiting the applicability of this technique in real-world complex underwater environments. To address these challenges, the paper proposed a high-fidelity 3D seafloor scene reconstruction method based on a cross-dimensional Gaussian normal transition field. In this method, a cross-dimensional mapping and normal transition system for Gaussian primitives was constructed, enabling fine-grained geometric modeling of complex structures. A Gaussian opacity-weighted filtering model was then presented to suppress reconstruction artifacts caused by medium-induced blurring effects. Experimental results on various underwater scenes demonstrate the proposed method’s capability to efficiently reconstruct complex underwater scenes and synthesize novel views.
-
图 1 跨维高斯法向跃迁场
Figure 1. Framework of cross-dimensional Gaussian normal transition field
图 5 不同方法对真实海底新视图合成的定性对比
Figure 5. Qualitative comparison of seafloor real scenario novel view synthesis using different methods
图 6 不同方法对真实海底三维几何重建的定性对比
Figure 6. Qualitative comparison of seafloor real scenario 3D geometric reconstruction using different methods
表 1 不透明度阈值调整分析
Table 1. Analysis of the opacity threshold adjustments
阈值 高斯数量 FPS PSNR SSIM 0.005 128.4×104 18.3 36.91 0.982 0.05 87.5×104 23.4 37.22 0.989 
下载: 导出CSV
表 2 海底场景数据集新视图合成定量对比
Table 2. Quantitative comparison of novel view synthesis for seafloor real scenario dataset
方法 PSNR SSIM LPIPS NeRF 30.152 0.910 0.181 Instant NGP 31.091 0.929 0.169 Mip-NeRF 32.471 0.941 0.147 Tri-MipRF 32.976 0.950 0.129 3DGS 36.249 0.977 0.051 2DGS 34.590 0.969 0.088 Mip-Splatting 36.254 0.981 0.055 SuGaR 32.175 0.937 0.151 GNTF 37.224 0.989 0.036
下载: 导出CSV
表 3 三维几何重建定量对比
Table 3. Quantitative comparison of 3D geometric reconstruction
方法 CD 运行时间/min SfM+PSR 7.89 43.5 3DGS+Marching cubes 2.63 23.9 GNTF+Marching cubes 1.98 19.2
下载: 导出CSV
表 4 消融实验定量结果对比
Table 4. Quantitative comparison of ablation experiments
方法 PSNR SSIM LPIPS CD GNTF 37.224 0.989 0.036 1.93 GNTF w/o CNT 36.933 0.971 0.052 2.33 GNTF w/o OWF 36.897 0.984 0.049 2.01
下载: 导出CSV
-
[1] Schonberger J L, Frahm J M. Structure-from-motion revisited[C]//Proceedings of the IEEE conference on computer vision and pattern recognition, 2016: 4104-4113. [2] Seitz S M, Curless B, Diebel J, et al. A comparison and evaluation of multi-view stereo reconstruction algorithms[C]//2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition(CVPR’06) IEEE, 2006: 519-528. [3] Dai P, Xu J, Xie W, et al. High-quality surface reconstruction using gaussian surfels[C]//ACM SIGGRAPH 2024, 2024: 1-11. [4] Mildenhall B, Srinivasan P P, Tancik M, et al. NeRF: representing scenes as neural radiance fields for view synthesis[J]. Communications of the ACM, 2021, 65(1): 99-106. [5] Barron J T, Mildenhall B, Tancik M, et al. Mip-NeRF: a multiscale representation for anti-aliasing neural radiance fields[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021: 5855-5864. [6] Barron J T, Mildenhall B, Verbin D, et al. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022: 5470-5479. [7] Hu W, Wang Y, Ma L, et al. Tri-MipRF: Tri-Mip representation for efficient anti-aliasing neural radiance fields[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023: 19774-19783. [8] Akkaynak D, Treibitz T. Sea-Thru: a method for removing water from underwater images[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019: 1682-1691. [9] Levy D, Peleg A, Pearl N, et al. SeaThru-NeRF: Neural radiance fields in scattering media[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023: 56-65. [10] Cong Y, Gu C, Zhang T, et al. Underwater robot sensing technology: A survey[J]. Fundamental Research, 2021, 1(3): 337-345. doi: 10.1016/j.fmre.2021.03.002 [11] Kerbl B, Kopanas G, Leimkühler T, et al. 3D Gaussian splatting for real-time radiance field rendering[J]. ACM Transactions on Graphics, 2023, 42(4): 1-14. [12] Yu Z, Chen A, Huang B, et al. Mip-Splatting: Alias-free 3D gaussian splatting[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024: 19447-19456. [13] Huang B, Yu Z, Chen A, et al. 2D gaussian splatting for geometrically accurate radiance fields[C]//ACM SIGGRAPH 2024, 2024: 1-11. [14] Hao Y, Yuan Y, Zhang H, et al. Underwater optical imaging: methods, applications and perspectives[J]. Remote Sensing, 2024, 16(20): 3773. doi: 10.3390/rs16203773 [15] Chu X, Chen H, Zou D, et al. A multimodal optical dataset for underwater image enhancement, detection, segmentation, and reconstruction[J]. Scientific Data, 2025, 12(1): 1554. doi: 10.1038/s41597-025-05797-w [16] Chu X. A multimodal optical dataset for underwater image enhancement, detection, segmentation, and reconstruction[DS/OL]. V2.0.Zenodo[2025-02-15]. https://doi.org/10.5281/zenodo.14865039.Chu X. A multimodal optical dataset for underwater image enhancement, detection, segmentation, and reconstruction[DS/OL]. V2.0.Zenodo[2025-02-15]. https://doi.org/10.5281/zenodo.14865039. [17] Yu Z, Sattler T, Geiger A. Gaussian opacity fields: efficient adaptive surface reconstruction in unbounded scenes[J]. ACM Transactions on Graphics, 2024, 43(6): 1-13. doi: 10.1145/3687937 [18] William E L, Harvey E C. Marching cubes: a high resolution 3D surface construction algorithm[J]. ACM SIGGRAPH Computer GraphicsS, 1987, 24(4): 163-169. [19] Müller T, Evans A, Schied C, et al. Instant neural graphics primitives with a multiresolution hash encoding[J]. ACM Transactions on Graphics, 2022, 41(4): 1-15. doi: 10.1145/3528223.3530127 [20] Guédon A, Lepetit V. Sugar: surface-aligned gaussian splat-ting for efficient 3D mesh reconstruction and high-quality mesh rendering[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024: 5354-5363. -
下载:
点击查看大图
计量
- 文章访问数: 448
- HTML全文浏览量: 265
- PDF下载量: 46
- 被引次数: 0
