ECNR: Efficient Compressive Neural Representation of Time-Varying Volumetric Datasets

TL;DR

ECNR proposes an efficient compressive neural representation for time-varying volumetric data by replacing a single large MLP with a multiscale, Laplacian-pyramid-based approach that fits local blocks with multiple small MLPs. A key novelty is block assignment uniformization via k-means to balance work across MLPs, together with a deep compression pipeline featuring block-guided pruning, global quantization, and entropy encoding, plus a lightweight CNN to mitigate boundary artifacts. Across four datasets, ECNR achieves higher data- and image-level fidelity (PSNR, LPIPS) and feature-level closeness (CD) at high compression rates, while delivering substantial speedups over neurcomp (encoding up to ~3.2× and decoding up to ~30×) and competitive results against SZ3 and TTHRESH. The work demonstrates a practical, scalable path for archiving large time-varying volumes, with room for future acceleration and extensions to multivariate data and streaming workflows.

Abstract

Due to its conceptual simplicity and generality, compressive neural representation has emerged as a promising alternative to traditional compression methods for managing massive volumetric datasets. The current practice of neural compression utilizes a single large multilayer perceptron (MLP) to encode the global volume, incurring slow training and inference. This paper presents an efficient compressive neural representation (ECNR) solution for time-varying data compression, utilizing the Laplacian pyramid for adaptive signal fitting. Following a multiscale structure, we leverage multiple small MLPs at each scale for fitting local content or residual blocks. By assigning similar blocks to the same MLP via size uniformization, we enable balanced parallelization among MLPs to significantly speed up training and inference. Working in concert with the multiscale structure, we tailor a deep compression strategy to compact the resulting model. We show the effectiveness of ECNR with multiple datasets and compare it with state-of-the-art compression methods (mainly SZ3, TTHRESH, and neurcomp). The results position ECNR as a promising solution for volumetric data compression.

Figures (19)

• Figure 1: Employing the Laplacian pyramid, ECNR decomposes a volume into blocks in terms of their low-resolution content (coarsest scale) and residuals (finer scales). A three-scale ($s=3$) example with the simplified 3D version (i.e., the temporal dimension is omitted) is sketched. A group of MLPs encodes all blocks at the same scale.
• Figure 1: Top to bottom: PSNR (dB), LPIPS, and CD values over timesteps. For CD, the chosen isovalues are reported in Table 2 in the paper.
• Figure 2: At each scale, the encoding target is split into equal-sized blocks, and only effective blocks with large residual values (shown in red bounding boxes) are processed.
• Figure 2: ECNR compression file storage percentages for different components.
• Figure 3: Average PSNR (dB) values across all timesteps under different CRs.