Interactive Volume Visualization via Multi-Resolution Hash Encoding Based Neural Representation.

Area of Science:

• Computer Graphics
• Scientific Visualization
• Machine Learning

Background:

• Implicit neural networks offer data compression for volume visualization.
• High training and inference costs limit current neural network applications to offline processing.

Purpose of the Study:

• To enable real-time direct ray tracing of volumetric neural representations.
• To overcome limitations of existing methods for interactive visualization of large-scale volume data.

Main Methods:

• Leveraging GPU tensor cores, CUDA, optimized volume rendering, and acceleration data structures.
• Implementing an out-of-core training strategy for extreme-scale data.
• Integrating training within a rendering loop to bypass pre-training.

Main Results:

• Achieved high-fidelity neural representations (PSNR > 30 dB) with data size reduction up to 1000x.
• Enabled real-time rendering and interactive visualization.
• Demonstrated terascale training on a single workstation GPU.

Conclusions:

• The novel approach significantly outperforms state-of-the-art methods in training time, reconstruction quality, and rendering performance.
• This method is ideal for applications requiring fast, accurate visualization of large-scale volume data.
• Real-time neural volume rendering is now feasible for interactive applications.