GIR: 3D Gaussian Inverse Rendering for Relightable Scene Factorization.

Area of Science:

• Computer Vision
• Computer Graphics
• Scientific Visualization

Background:

• Inverse rendering aims to reconstruct scene properties from images.
• Existing methods often struggle with complex lighting and real-time performance.
• 3D Gaussian representations offer a powerful new paradigm for scene representation.

Purpose of the Study:

• To develop a 3D Gaussian Inverse Rendering (GIR) method.
• To enable efficient factorization of scenes into material, light, and geometry.
• To achieve state-of-the-art performance in relighting and novel view synthesis.

Main Methods:

• Utilized 3D Gaussian representations for scene factorization.
• Computed surface normals using shortest eigenvectors with directional masking.
• Implemented voxel-based indirect illumination tracing for multi-bounce light transport.
• Employed a lightweight convolutional network for environmental map representation.

Main Results:

• Achieved accurate normal estimation without external supervision.
• Successfully disentangled secondary illumination for realistic light transport.
• Demonstrated state-of-the-art performance in relighting and novel view synthesis.
• Realized real-time rendering capabilities.

Conclusions:

• The proposed 3D Gaussian Inverse Rendering (GIR) method is effective and broadly applicable.
• The method shows significant potential for real-time interactive graphics applications.
• This work advances the field of inverse rendering and real-time graphics.