Geometry-Semantics Co-Regularization for Gaussian Splatting in Indoor Reconstruction

Area of Science:

• Computer Vision
• Computer Graphics
• Robotics

Background:

• 3D Gaussian Splatting (3DGS) shows promise for indoor scene reconstruction.
• Existing 3DGS methods often produce artifacts due to weak 2D monocular priors.
• Robust indoor scene reconstruction is crucial for AR/VR, robotics, and autonomous systems.

Purpose of the Study:

• To develop a geometry-semantics co-regularization framework for 3DGS.
• To enhance the geometric accuracy and semantic consistency of reconstructed 3D scenes.
• To mitigate artifacts like holes and unnatural protrusions in 3D reconstructions.

Main Methods:

• Proposed a geometry-semantics co-regularization framework for joint optimization within 3DGS.
• Introduced a dual implicit architecture integrating Implicit Moving Least Squares (IMLS) and Signed Distance Functions (SDFs) for geometric optimization.
• Developed a view-dependent normal consistency constraint and multi-view semantic consistency supervision.

Main Results:

• Achieved state-of-the-art geometric reconstruction quality on Replica and MuSHRoom datasets.
• Demonstrated improved rendering fidelity compared to existing methods.
• Significantly reduced artifacts such as holes and unnatural protrusions.

Conclusions:

• The proposed geometry-semantics co-regularization framework effectively improves 3DGS for indoor scene reconstruction.
• The integration of dual implicit representations and semantic consistency enhances both geometric and semantic accuracy.
• The method offers a robust solution for high-fidelity 3D scene reconstruction with reduced artifacts.