Make the Fastest Faster: Importance Mask Synthesis for Interactive Volume Visualization Using Reconstruction Neural...

Area of Science:

• Computer Graphics
• Scientific Visualization
• Deep Learning

Background:

• Visualizing large-scale, high-resolution volumetric data is computationally intensive.
• Current deep learning methods reduce latency but still require significant pixel rendering.
• Existing approaches render pixels based on regular or irregular sampling patterns.

Purpose of the Study:

• To minimize pixel rendering by directly synthesizing important regions.
• To develop a unified framework for various inpainting methods.
• To improve rendering latency in scientific volume visualization.

Main Methods:

• Introduced Importance Mask Learning (IML) and Synthesis (IMS) networks.
• Developed differentiable compaction/decompaction layers for a unified framework.
• Jointly considered dataset, user behavior, and reconstruction networks.

Main Results:

• Significantly improved rendering latency for state-of-the-art volume visualization methods.
• Achieved further reductions in rendering time compared to existing deep learning methods.
• Enabled direct optimization of pre-trained reconstruction networks without extensive retraining.

Conclusions:

• IML and IMS networks offer a novel approach to accelerate scientific volume visualization.
• The unified framework enhances compatibility with diverse inpainting techniques.
• This method provides substantial performance gains for large volumetric datasets.