Learning to predict perceptual visibility of rendering deterioration in computer games

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Computer gamers manually adjust rendering settings (texture resolution, shadow maps, anti-aliasing) to balance performance and visual fidelity.
• The complex interplay of these settings makes manual optimization challenging, especially on lower-end hardware.
• Automated methods are needed to simplify the selection of optimal rendering configurations.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for automated optimization of game rendering settings.
• To enable efficient identification and classification of visual artifacts caused by reduced rendering quality.
• To assist players in selecting configurations that minimize artifact visibility and maximize performance.

Main Methods:

• Training a proprietary CNN to compare high-quality reference images with reduced-quality rendered images.
• The CNN identifies and classifies visual artifacts and assesses their perceptual impact.
• The network was trained on diverse game engine scenes with manually annotated quality-diminished regions.

Main Results:

• The developed CNN effectively identifies and classifies visual artifacts in rendered images.
• Implementation in a prototype OpenGL renderer allowed evaluation of anti-aliasing settings.
• User studies confirmed the CNN's superior ability to discern artifact visibility over existing metrics.

Conclusions:

• The CNN provides an effective automated solution for optimizing game rendering settings.
• This approach simplifies the process of achieving optimal visual quality and performance.
• The method surpasses traditional image quality metrics in artifact detection efficacy.