3D Gaussian Splatting with Depth and Normal Priors for Improved Indoor Scene Reconstruction
Core Concepts
Incorporating depth and normal priors into the optimization of 3D Gaussian splatting improves the photorealism and geometric accuracy of indoor scene reconstructions, enabling efficient mesh extraction directly from the optimized Gaussian representation.
Abstract
The paper presents a method called "DN-Splatter" that extends 3D Gaussian splatting, a differentiable rendering technique, by incorporating depth and normal priors to improve the reconstruction quality of indoor scenes.
Key highlights:
Depth regularization is applied to the optimization of 3D Gaussian splatting, using sensor depth data or monocular depth estimates, to better align the Gaussian positions with the true scene geometry.
Smoothness constraints are enforced on the rendered depth maps to ensure smooth depth transitions.
Normal priors obtained from monocular normal estimation networks are used to align the orientations of the 3D Gaussians with the scene geometry.
The regularized Gaussian scene representation is then used to directly extract meshes using Poisson surface reconstruction, without the need for additional optimization or refinement stages.
Experiments on indoor datasets show that the proposed depth and normal regularization strategy significantly improves the quality of the reconstructed meshes compared to baseline 3D Gaussian splatting and other neural implicit reconstruction methods.
DN-Splatter
Stats
3D Gaussian splatting can represent a scene with millions of differentiable 3D Gaussian primitives with optimizable geometric and appearance properties.
Depth regularization is applied using a gradient-aware logarithmic depth loss and a total variation loss on the rendered depth maps.
Normal priors are obtained from a pre-trained monocular normal estimation network and used to align the orientations of the 3D Gaussians.
Quotes
"Incorporating depth and normal priors into the optimization of 3D Gaussian splatting improves the photorealism and geometric accuracy of indoor scene reconstructions, enabling efficient mesh extraction directly from the optimized Gaussian representation."
"We regularize the optimization of 3D Gaussian splatting in indoor scenes with depth and smoothness constraints enhancing novel view synthesis results whilst respecting the captured scene geometry better."
"We use monocular normal priors to align Gaussians with the scene geometry and show how this results in more accurate geometry reconstructions."
How could the proposed depth and normal regularization strategy be extended to handle more challenging capture conditions, such as motion blur and other artifacts, in sparse view settings
What other types of geometric priors or scene representations could be integrated with 3D Gaussian splatting to further improve the quality and efficiency of the reconstructions
How could the concurrent optimization of the Gaussian scene representation and the extracted mesh be explored to achieve even smoother and more accurate reconstructions
3D Gaussian Splatting with Depth and Normal Priors for Improved Indoor Scene Reconstruction
DN-Splatter
How could the proposed depth and normal regularization strategy be extended to handle more challenging capture conditions, such as motion blur and other artifacts, in sparse view settings
What other types of geometric priors or scene representations could be integrated with 3D Gaussian splatting to further improve the quality and efficiency of the reconstructions
How could the concurrent optimization of the Gaussian scene representation and the extracted mesh be explored to achieve even smoother and more accurate reconstructions