GMLight: Lighting Estimation via Geometric Distribution Approximation.

Area of Science:

• Computer Vision
• Computer Graphics
• Image Processing

Background:

• Estimating scene illumination from single images is crucial but difficult.
• Current methods for lighting estimation often lack accuracy and generalizability.
• These methods include regressing illumination parameters or directly generating illumination maps.

Purpose of the Study:

• To introduce Geometric Mover's Light (GMLight), a novel framework for accurate scene illumination estimation.
• To address the limitations of existing methods in terms of accuracy and generalization.
• To enable high-fidelity relighting for 3D object insertion.

Main Methods:

• GMLight utilizes a regression network and a generative projector for illumination estimation.
• Illumination scenes are parameterized by geometric light distribution, intensity, ambient term, and depth.
• A novel geometric mover's loss, inspired by earth mover's distance, guides parameter regression.

Main Results:

• The proposed geometric mover's loss effectively guides the accurate regression of light distribution parameters.
• The generative projector synthesizes panoramic illumination maps with realistic appearance and high-frequency details.
• GMLight demonstrates accurate illumination estimation and superior fidelity in relighting tasks.

Conclusions:

• GMLight offers a robust and accurate solution for single-image scene illumination estimation.
• The framework shows significant improvements in relighting applications, particularly for 3D object insertion.
• The novel geometric mover's loss is key to achieving precise light distribution parameter estimation.