A Machine-Learning-Driven Sky Model.

Area of Science:

• Computer Graphics
• Machine Learning
• Virtual Environments

Background:

• Sky illumination is crucial for realistic virtual environments.
• Existing methods for representing sky lighting can be memory-intensive.
• Creating smooth lighting transitions often requires numerous pre-captured or complex models.

Purpose of the Study:

• To develop a compact machine learning-based representation of sky illumination.
• To enable efficient approximation of captured and analytical sky lighting.
• To facilitate smooth sky lighting transitions using minimal data.

Main Methods:

• A machine learning approach was used to learn a compact representation of sky illumination.
• The method was applied to both analytical sky models and captured environment maps.
• The representation was evaluated for its memory efficiency and accuracy.

Main Results:

• The proposed approach significantly reduces memory costs for representing sky illumination.
• It accurately approximates captured lighting, achieving results close to ground truth.
• Smooth transitions between different times of day were successfully generated from a small set of environment maps.

Conclusions:

• The machine learning method provides an efficient and accurate way to represent sky illumination in virtual environments.
• The approach offers a low runtime overhead, suitable for both offline and real-time rendering.
• This technique enables more dynamic and memory-efficient virtual lighting solutions.