Parallel Compositing of Volumetric Depth Images for Interactive Visualization of Distributed Volumes at High Frame Rates

TL;DR

The paper addresses the challenge of interactive visualization of large distributed volumes by introducing Volumetric Depth Images (VDIs) as view-dependent, compact proxies for volume data. It presents a scalable, sort-last distributed generation and parallel compositing algorithm that creates sub-VDIs per processing element and coherently merges them into a final VDI, with a compact in-memory layout to reduce communication overhead. The approach supports non-convex domain decompositions and employs approximate per-ray gamma optimization to maximize detail under a subsegment budget, enabling high frame-rate remote visualization with strong fidelity to direct volume rendering. Timings, fidelity metrics (SSIM/PSNR), and extensive benchmarks demonstrate substantial speedups over distributed DVR, particularly at near viewpoints, while maintaining high visual quality and enabling in-situ or remote interaction for very large datasets. The work provides open-source tooling and practical guidance for deploying VDIs in distributed visualization pipelines, highlighting both potential impact and current limitations (e.g., transfer-function interactivity and lighting effects).

Abstract

We present a parallel compositing algorithm for Volumetric Depth Images (VDIs) of large three-dimensional volume data. Large distributed volume data are routinely produced in both numerical simulations and experiments, yet it remains challenging to visualize them at smooth, interactive frame rates. VDIs are view-dependent piecewise constant representations of volume data that offer a potential solution. They are more compact and less expensive to render than the original data. So far, however, there is no method for generating VDIs from distributed data. We propose an algorithm that enables this by sort-last parallel generation and compositing of VDIs with automatically chosen content-adaptive parameters. The resulting composited VDI can then be streamed for remote display, providing responsive visualization of large, distributed volume data.

Figures (13)

• Figure 1: A Volumetric Depth Image (VDI) Frey is generated by casting rays through the volume and grouping samples (depicted by colored circles along the ray) of similar color and opacity, generating a list $\mathbb{L}_{i}$ of up to $\mathbb{N_{S}}$=3 (in the example of the figure) supersegments $\mathbb{S}^{i}_{j}$ per ray. Each $\mathbb{S}^{i}_{j}$ stores its front and back face, $f(\mathbb{S}^{i}_{j})$ and $b(\mathbb{S}^{i}_{j})$, along with color and opacity accumulated in-between (Fig. \ref{['fig:dense']}a). Hollow circles represent samples in empty regions. Volume data may be divided among multiple Processing Elements (PE) in a computer cluster (background gray levels).
• Figure 2: Representing the VDI generated in Fig. \ref{['fig:generation']} in memory. a) All 18 $\mathbb{S}^{}_{}$ are stored in memory. b) Prefix sum evaluated on (a), which is used to generate the compact representation of the VDI shown in c), storing only the 11 non-empty $\mathbb{S}^{}_{}$.
• Figure 3: The sub-VDIs generated during Phase 1 and the lists composited during Phase 2 for the VDI in Fig. \ref{['fig:generation']}. Panels c) and d) show sub-$\mathbb{S}^{}_{}$ sorted by their position along the ray. Flat outlines represent sub-$\mathbb{S}^{}_{}$ received from PE 1, while dotted outlines represent sub-$\mathbb{S}^{}_{}$ received from PE 2. These are then composited, producing a maximum of $\mathbb{N_{S}}$$=3$$\mathbb{S}^{}_{}$ per list, as shown in Fig. \ref{['fig:generation']}.
• Figure 4: SSIM (mean $\pm$ min-max range across four viewpoints of VDI generation) with respect to ground truth DVR for the rendering of VDIs generated using varying numbers of Nvidia A100 GPUs. VDIs generated on 8, 16, 24 and, 32 GPUs are composited using the presented compositing algorithm (Sec. \ref{['sec:compVDI']}), while VDIs generated on 1 GPU do not undergo compositing. All VDIs are of resolution $\mathbb{N_{L}}$=1920$\times$1080 with $\mathbb{N_{S}}$=25 for three datasets (symbols, top legend) and three different $\mathrm{V}_\mathrm{N}$ (panels).
• Figure 5: Visual illustration of VDI rendering quality. VDIs generated on 32 GPUs are rendered at 20° from the viewpoint of generation. SSIM values computed w.r.t. ground-truth direct volume rendering (DVR) at the same viewpoint (see Supplement for images).