A Study of Data Augmentation for Learning-Driven Scientific Visualization.

Area of Science:

• Scientific Visualization
• Deep Learning
• Data Augmentation

Background:

• Deep learning requires extensive data, which is scarce in scientific visualization due to high computational costs.
• Data augmentation is a key technique to address data sparsity and improve model performance.

Purpose of the Study:

• To comprehensively study nine data augmentation techniques for scientific visualization tasks.
• To evaluate their effectiveness on spatial super-resolution and ambient occlusion prediction.

Main Methods:

• Compared noise injection, interpolation, scale, flip, rotation, variational auto-encoder, generative adversarial network, diffusion model, and implicit neural representation.
• Assessed data quality, rendering fidelity, optimization time, and memory consumption across diverse scientific datasets.
• Investigated the impact of augmentation method, quantity, and diversity on various deep learning models.

Main Results:

• Increasing the quantity and single-domain diversity of augmented data significantly boosts model performance.
• The method and cross-domain diversity of augmented data showed less impact on performance.
• Performance gains were observed for spatial super-resolution and ambient occlusion prediction tasks.

Conclusions:

• Data augmentation is crucial for overcoming data limitations in scientific visualization deep learning.
• Optimizing the quantity and single-domain diversity of augmented data is key for performance enhancement.
• Future research should explore novel augmentation strategies and their impact on diverse scientific visualization challenges.