Efficient representation of 3D spatial data for defense-related applications

TL;DR

This paper analyzes how 3D spatial data can be efficiently represented for defense applications by comparing traditional geometry-centric representations (points, voxels, meshes) with neural and explicit representations (NeRFs, Instant-NGP, Gaussian Splatting). It finds a clear trade-off: traditional models provide robust geometric fidelity for tasks like line-of-sight analysis and physics simulations, while modern neural methods deliver high-fidelity visuals but often falter on geometric reliability and scalability. To address these gaps, the authors advocate a hybrid architecture that couples a geometric scaffold (mesh) for accurate geometry with a surface-focused neural representation (3DGS) for visual detail within a hierarchical scene structure (e.g., BVH). This hybrid approach aims to deliver scalable, photorealistic, and geometrically grounded 3D battlespace representations suitable for decision-making, visualization, and simulation in defense contexts.

Abstract

Geospatial sensor data is essential for modern defense and security, offering indispensable 3D information for situational awareness. This data, gathered from sources like lidar sensors and optical cameras, allows for the creation of detailed models of operational environments. In this paper, we provide a comparative analysis of traditional representation methods, such as point clouds, voxel grids, and triangle meshes, alongside modern neural and implicit techniques like Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS). Our evaluation reveals a fundamental trade-off: traditional models offer robust geometric accuracy ideal for functional tasks like line-of-sight analysis and physics simulations, while modern methods excel at producing high-fidelity, photorealistic visuals but often lack geometric reliability. Based on these findings, we conclude that a hybrid approach is the most promising path forward. We propose a system architecture that combines a traditional mesh scaffold for geometric integrity with a neural representation like 3DGS for visual detail, managed within a hierarchical scene structure to ensure scalability and performance.

Figures (15)

• Figure 1: Screenshot from the popular geospatial information software WinTAKWintak (left), and ambient-light lidar-scan of the same area (right). A 3D representation can provide richer information and enhance decision-making.
• Figure 2: Generalized steps of a sensor-to-representation pipeline. Raw sensor data is collected, fused, then an aggregate model is formed, which is then queried and visualized by the end-user application.
• Figure 3: Overview of the first two stages. Derivative data combines and filters raw data from multiple sensors.
• Figure 4: Examples of common use cases for a 3D representation model.
• Figure 5: An example of a multi-layer point-cloud captured by a lidar sensor modissa, consising of points (top-left), pulse-intensity (top-right), ambient light (bottom-left) and RGB colors captured by a complementary camera rig (bottom-right).