DCINR: A Divide-and-Conquer Implicit Neural Representation for Compressing Time-Varying Volumetric Data in Hours

Area of Science:

• Computer Vision
• Data Compression
• Machine Learning

Background:

• Implicit neural representations (INRs) are effective for compressing time-varying volumetric data.
• Current INR optimization is computationally intensive, often requiring days or weeks for completion.

Purpose of the Study:

• To introduce a novel method, divide-and-conquer INR (DCINR), for significantly accelerating the compression of time-varying volumetric data.
• To improve compression ratios and visual fidelity compared to existing methods.

Main Methods:

• The dataset is divided into non-overlapping blocks.
• A block selection strategy is employed to remove redundant blocks, reducing computational cost.
• Each selected block is modeled by a tiny INR, with size adapted to information richness, determined by maximizing average network capacity.

Main Results:

• DCINR reduces compression time for time-varying volumetric data to hours.
• Achieves superior compression ratios (thousands to tens of thousands) and visual fidelity over learning-based and lossy compression methods.
• Compression time is comparable to lossy compressors while preserving high-quality features.

Conclusions:

• DCINR offers a highly efficient and effective solution for compressing time-varying volumetric data.
• The method achieves extreme compression ratios with excellent visual quality and reduced processing time.
• DCINR represents a significant advancement in volumetric data compression technology.