Volumetric Isosurface Rendering with Deep Learning-Based Super-Resolution.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Scientific Visualization

Background:

• Accurate isosurface rendering in volumetric data demands extensive sampling.
• Reducing data requirements is a key challenge in volume rendering research.
• Deep learning offers potential for inferring missing data samples, akin to image super-resolution.

Purpose of the Study:

• Investigate deep learning for upscaling low-resolution isosurface sampling to higher resolutions.
• Reconstruct spatial detail and shading for enhanced isosurface visualization.
• Improve frame-to-frame coherence in dynamic rendering scenarios.

Main Methods:

• Developed a fully convolutional neural network (CNN) for learning latent representations.
• Generated smooth, edge-aware depth, normal fields, and ambient occlusions from low-resolution inputs.
• Incorporated frame-to-frame motion loss to account for temporal variations during training.

Main Results:

• The CNN effectively reconstructs spatial detail and shading for unseen isosurfaces.
• Upscaled results demonstrate superior quality compared to traditional bi-linear and cubic methods.
• Frame-to-frame motion loss significantly improves temporal coherence in dynamic scenes.

Conclusions:

• Deep learning, specifically CNNs with temporal loss, offers a powerful approach for efficient and high-quality isosurface rendering.
• This method holds promise for applications like remote visualization and foveated rendering.
• Training on similar data further enhances the network's performance for specific isosurface types.