International Open Access Journal Platform

基于经典方法与新兴融合技术的三维建模比较分析

Comparative Analysis of Classical Methods and Emerging Fusion Technologies for 3D Modeling

方子威, 文静, 黄于蒙, 高昕

Progress in Social Sciences / 2026,8(7): 638-647 / 2026-08-04 13 8

View Download PDF

• Information:

  江苏师范大学物理与光电工程学院，徐州

• Keywords:
  3D reconstruction; Neural radiance fields; 3D Gaussian splatting; Multi-view stereo; Novel view synthesis
  三维重建; 神经辐射场; 3D高斯泼溅; 多视图立体; 新视角合成
• Abstract: Reconstructing 3D structures from 2D images has long been a focal research topic in computer vision and computer graphics. Four representative technical routes named as structure from Motion (SfM), Multi-View Stereo (MVS), Neural Radiance Fields (NeRF), and 3D Gaussian Splatting (3DGS) are selected by this paper. With respect to the public datasets of two benchmarks, Mip-NeRF 360 and Tanks & Temples, quantitative evaluations on their reconstruction accuracy, rendering efficiency, resource consumption, and scene applicability are conducted. Experimental results show that 3DGS achieves a leading real-time rasterization speed of 134~197 FPS, which is approximately 1000~3000 times faster than Mip-NeRF 360. The training time of 3DGS is reduced to 35~45 minutes, while its PSNR (27.21 dB) is nearly comparable to that of Mip-NeRF 360 (27.69 dB). The classic workflow still maintains advantages in geometric accuracy and editability, while neural representations demonstrate remarkable superiority for novel view synthesis under sparse-view settings. 从二维图像恢复三维结构的问题，长期以来备受计算机视觉与图形学领域关注。本文选取运动恢复结构（SfM）、多视图立体（MVS）、神经辐射场（NeRF）与3D高斯泼溅（3DGS）四条路线，以Mip-NeRF360和Tanks&Temples两个基准上的公开数据集为基础，定量评估其重建精度、渲染效率、资源消耗与场景适用性。数据表明，3DGS以134～197FPS的实时光栅化速率领先，约为Mip-NeRF360的1000～3000倍，训练压缩至35～45分钟，PSNR（27.21dB）与Mip-NeRF360（27.69dB）基本持平。经典工作流程在几何精度和可编辑性上依然领先。在稀疏视角下的新视角合成中，神经表征优势明显。
• DOI: 10.35534/pss.0807110 (registering DOI)
• Cite: 方子威，文静，黄于蒙，等．基于经典方法与新兴融合技术的三维建模比较分析［J］．社会科学进展，2026，8（7）：638-647.