17 Jul 2024 | Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Juho Kannala, Esa Rahtu
**DN-Splatter: Depth and Normal Priors for Gaussian Splatting and Meshing**
This paper addresses the challenge of high-fidelity 3D reconstruction of common indoor scenes, crucial for VR and AR applications. It introduces DN-Splatter, a method that extends 3D Gaussian splatting with depth and normal priors to improve geometric constraints during optimization. The key contributions include:
1. **Edge-Aware Depth Loss**: A gradient-aware depth loss is proposed to regularize the position of Gaussians, enhancing depth estimation and novel view synthesis.
2. **Monocular Normal Regularization**: Normals are estimated from Gaussians and used to align them with the true scene geometry, improving surface reconstruction.
3. **Mesh Extraction**: The optimized Gaussian scene is directly used for mesh extraction using Poisson surface reconstruction, resulting in smoother and more accurate reconstructions.
The method is evaluated on challenging indoor datasets, demonstrating superior performance in novel view synthesis, depth estimation, and mesh quality compared to state-of-the-art methods. The code is available at <https://github.com/maturk/dn-splatter>.**DN-Splatter: Depth and Normal Priors for Gaussian Splatting and Meshing**
This paper addresses the challenge of high-fidelity 3D reconstruction of common indoor scenes, crucial for VR and AR applications. It introduces DN-Splatter, a method that extends 3D Gaussian splatting with depth and normal priors to improve geometric constraints during optimization. The key contributions include:
1. **Edge-Aware Depth Loss**: A gradient-aware depth loss is proposed to regularize the position of Gaussians, enhancing depth estimation and novel view synthesis.
2. **Monocular Normal Regularization**: Normals are estimated from Gaussians and used to align them with the true scene geometry, improving surface reconstruction.
3. **Mesh Extraction**: The optimized Gaussian scene is directly used for mesh extraction using Poisson surface reconstruction, resulting in smoother and more accurate reconstructions.
The method is evaluated on challenging indoor datasets, demonstrating superior performance in novel view synthesis, depth estimation, and mesh quality compared to state-of-the-art methods. The code is available at <https://github.com/maturk/dn-splatter>.