Grunnleggende konsepter
GeoGaussian enhances 3D Gaussian rendering quality by preserving accurate geometry and introducing novel densification strategies.
Sammendrag
The article introduces GeoGaussian, a method focusing on improving the rendering quality of 3D Gaussians in non-textured regions. It proposes geometry-aware strategies for initialization and densification to enhance rendering performance. The method is evaluated on public datasets, showcasing superior results compared to state-of-the-art approaches.
- Introduction to Neural Radiance Fields (NeRF) and Gaussian Splatting.
- Challenges in maintaining scene geometry during Gaussian Splatting optimization.
- Proposal of GeoGaussian approach emphasizing geometry preservation.
- Detailed methodology including Gaussian initialization, densification, and view-dependent optimization.
- Comparison with existing methods on public datasets showcasing improved rendering quality.
- Ablation studies on sparse training views and evaluation on different datasets.
Statistikk
During the first 15K iterations, our method achieves significantly better results than 3DGS.
Our model trained on ICL-O3 (10%) achieves a PSNR of 21.42, outperforming models of 3DGS and LightGS.
Sitater
"Our proposed pipeline achieves state-of-the-art performance in novel view synthesis and geometric reconstruction."
"GeoGaussian preserves the reasonable geometry of environments compared to 3DGS."