Authors: Xiong Jinlin, Li Can, Shen Jiawei, Qi Zhigang, Sun Lei, Zhao Dongyang
Organizations: Jinlin Xiong, Can Li, Jiawei Shen and Lei Sun are with Tianjin Key Laboratory of Intelligent Robotics, Nankai University, Tianjin 300350, China (tjKLIR). · Dongyang Zhao and Zhigang Qi are with Beijing Institute of Computer Application, Beijing 100081, China.
High-fidelity three-dimensional (3D) reconstruction is essential for robotics and simulation. While Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) achieve impressive rendering quality, their reliance on time-consuming per-scene optimization limits real-time deployment. Emerging feed-forward Gaussian splatting methods improve efficiency but often lack explicit, manifold geometry required for direct simulation. To address these limitations, we propose a feed-forward framework for triangle primitive generation that directly predicts continuous triangle surfaces from calibrated multi-view images. Our method produces simulation-ready models in a single forward pass, obviating the need for per-scene optimization or post-processing. We introduce a pixel-aligned triangle generation module and incorporate relative 3D point cloud supervision to enhance geometric learning stability and consistency. Experiments demonstrate that our method achieves efficient reconstruction while maintaining seamless compatibility with standard graphics and robotic simulators.
Figures & tables
Fig. 2 : Overview of the proposed feed-forward triangular surface reconstruction network. Multi-view images are processed by a Multi-View Depth Estimation module to obtain fused features enriched with depth information. The fused features are used to predict depth maps and back-project an initial 3D point cloud, while a 2D U-Net with a triangle head decodes additional vertex attributes (opacity and spherical harmonics color). A surface generation module infers face connectivity to produce the final triangular surface. Differentiable rasterization enables photometric supervision, and external 3D point cloud supervision provides explicit geometric constraints during training.
Fig. 3 : Qualitative comparison of reconstruction quality between our method and optimization-based triangle rasterization methods.
Fig. 4 : Qualitative comparison of 3D spatial consistency between our method and feed-forward Gaussian Splatting methods.
algorithm
PSNR ↑
SSIM ↑
LPIPS ↓
iter. ↓
con.
3DGS
18.35
0.601
0.350
30k
×
2DGS
16.17
0.526
0.419
30k
×
trianglesplatting
18.53
0.683
0.279
30k
×
meshsplatting
19.78
0.685
0.340
30k
✓
FTSplat(Ours)
20.39
0.707
0.257
1
✓
TABLE I : Quantitative comparison of reconstruction quality with optimization-based methods. (iter.: number of iterations; con.: connectivity, i.e. , whether the generated representation is geometrically connected.)
algorithm
PSNR ↑
SSIM ↑
LPIPS ↓
Mvsplat
27.03
0.891
0.106
Depthsplat
27.61
0.903
0.099
FTSplat(Ours)
20.39
0.707
0.257
TABLE II : Quantitative comparison of reconstruction quality with feed-forward Gaussian Splatting methods.
algorithm
PSNR ↑
SSIM ↑
LPIPS ↓
w/o
13.06
0.401
0.509
w/ vggt
20.07
0.692
0.275
w/ da3
20.39
0.707
0.257
TABLE III : Quantitative ablation study on relative 3D point cloud supervision.(w/o: without point cloud supervision; w/: point cloud signals provided by VGGT or Depth Anything 3.)
Fig. 5 : Qualitative comparison of rendering quality under different ablation settings
Fig. 6 : Qualitative comparison of 3D structure reconstruction under different ablation settings