Deep-Saliency Foveated Ray Tracing For Real-time VR Rendering.

Area of Science:

• Computer Graphics
• Human-Computer Interaction
• Virtual Reality

Background:

• High-resolution, high-refresh-rate VR demands significant real-time rendering resources.
• Conventional foveated rendering uses simplified models, neglecting attentional cues and causing peripheral artifacts.

Purpose of the Study:

• To develop an advanced foveated rendering framework that overcomes limitations of current methods.
• To improve perceptual quality and reduce artifacts in VR rendering.

Main Methods:

• Introduced a motion-aware foveation model to capture temporal dynamics.
• Employed a lightweight convolutional neural network to predict saliency maps from eye-gaze data.
• Integrated saliency maps to guide adaptive path tracing and filtering.

Main Results:

• The neural saliency-driven approach significantly reduces rendering artifacts.
• Perceptual quality is improved compared to previous foveated rendering techniques.
• Real-time performance is maintained.

Conclusions:

• Neural saliency-driven foveated rendering offers a promising solution for perceptually optimized VR.
• Accurate attentional modeling enhances rendering efficiency and visual fidelity.
• This framework advances real-time rendering for immersive applications.