F-Hash: Feature-Based Hash Design for Time-Varying Volume Visualization via Multi-Resolution Tesseract Encoding.

Area of Science:

• Computer Graphics
• Scientific Visualization
• Machine Learning

Background:

• Interactive time-varying volume visualization faces challenges with large datasets and complex spatiotemporal features.
• Implicit Neural Representations (INRs) are used for compression and super-resolution but suffer from slow convergence during training.
• Existing input encoding methods struggle with the efficiency required for large-scale time-varying volumetric data.

Purpose of the Study:

• To develop a novel encoding architecture that accelerates the convergence of Implicit Neural Representations for time-varying volumetric data.
• To improve the efficiency of interactive visualization and feature tracking for complex dynamic datasets.
• To provide a unified encoding solution for various feature detection methods.

Main Methods:

• Proposed F-Hash, a feature-based multi-resolution Tesseract encoding architecture.
• Incorporated multi-level collision-free hash functions for mapping dynamic 4D multi-resolution embedding grids.
• Developed an adaptive ray marching algorithm for optimized sample streaming during rendering.

Main Results:

• F-Hash demonstrated state-of-the-art convergence speed in training time-varying volumetric datasets.
• Achieved high encoding capacity with compact parameters using collision-free hash functions.
• The proposed method is agnostic to specific feature detection techniques, offering a unified approach.

Conclusions:

• F-Hash significantly enhances training convergence speed for time-varying neural representations.
• The method offers an efficient and unified solution for interactive volume visualization and feature evolution.
• Optimized rendering further improves the performance of visualizing dynamic volumetric data.