GeoTexDensifier: Geometry-Texture-Aware Densification for High-Quality Photorealistic 3D Gaussian Splatting.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• 3D Gaussian Splatting (3DGS) offers photorealistic rendering for applications like VR and 3D simulation.
• High-quality 3DGS reconstruction requires optimal splat distribution to capture geometric and texture details, which is challenging.
• Existing methods struggle with achieving sufficient splat density and accurate placement, especially in complex scenes.

Purpose of the Study:

• To introduce GeoTexDensifier, a novel strategy for reconstructing high-quality 3D Gaussian splats.
• To enhance compliance of splats with scene geometry and texture richness.
• To improve novel view synthesis and overall 3D model photorealism.

Main Methods:

• A texture-aware densification method increases splat density in highly textured regions.
• A geometry-aware splitting strategy utilizes depth and normal priors for precise splat placement.
• Validation of Depth Ratio Change filters noisy splats, guided by relative monocular depth priors.

Main Results:

• GeoTexDensifier produces denser splat distributions in textured areas and sparser ones in low-texture regions.
• Geometry-aware validation effectively reduces the impact of scattered Gaussians, improving rendering quality.
• Experiments show GeoTexDensifier generates more photorealistic 3DGS models compared to state-of-the-art methods.

Conclusions:

• GeoTexDensifier significantly enhances the quality of 3D Gaussian Splatting reconstructions.
• The combined geometry- and texture-aware strategies lead to superior novel view synthesis.
• This method offers a robust solution for creating high-fidelity 3D models from various datasets.