Evaluation of Novel View Synthesis in the Context of Radiometric Drone Imagery

Recent advancements in neural scene representations, specifically Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), have revolutionized novel-view synthesis. However, these methods are predominantly optimized for and evaluated on ground-based, visible-spectrum (RGB) data, leaving a significant gap in their application to aerial, radiometric thermal imagery, which is critical for industrial inspection, search-and-rescue, and environmental monitoring. In this work, we systematically investigate the applicability of these state-of-the-art paradigms to radiometric thermal imagery acquired from airborne drone platforms. We introduce a novel, publicly available multimodal dataset captured using a DJI M30T system, comprising synchronized RGB and radiometric thermal frames of a building. We conduct a comprehensive evaluation comparing specialized thermal approaches (ThermalNeRF, ThermoNeRF, Thermal3DGS) against general-purpose methods (nerfacto, gsplat). Our assessment utilizes a suite of quantitative metrics (PSNR, SSIM, MAE, LPIPS, and DISTS) complemented by qualitative visual analysis. Results indicate that Thermal3DGS achieves state-of-the-art performance in the thermal domain (PSNR 22.99, SSIM 0.845), effectively mitigating artifacts common in low-texture thermal data. Conversely, gsplat demonstrates superior RGB synthesis and competitive thermal performance, suggesting that general-purpose splatting representations are robust enough for cross-spectral applications. This work bridges the gap between aerial radiometric sensing and neural rendering, demonstrating that off-the-shelf drone thermography can be utilized for high-fidelity 3D thermal reconstruction with minimal adaptation.