Neural Radiance Fields and Reconstruction Fidelity in Urban Heritage Scans: Causal Modeling

Abstract

The rapid advancement of neural volume rendering has fundamentally transformed the landscape of three-dimensional digital preservation, particularly concerning complex urban heritage sites. Neural radiance fields have emerged as a dominant paradigm due to their ability to synthesize highly photorealistic novel views from a sparse set of two-dimensional images. However, the exact causal mechanisms determining the fidelity of these reconstructions remain poorly understood, as most existing literature relies heavily on associational observations rather than rigorous causal frameworks. This paper bridges this critical gap by introducing a structural causal model designed specifically to untangle the complex web of interactions between environmental conditions, geometric complexity, camera acquisition parameters, and the ultimate reconstruction fidelity of urban heritage scans. By employing Pearlian causal inference methodologies, this research rigorously identifies and isolates confounding variables such as ambient illumination fluctuations and structural occlusion that routinely corrupt observational data. Extensive interventional simulations are performed across a diverse portfolio of historical architectural datasets to quantify the direct and indirect causal effects of these variables on standard rendering metrics. The findings presented herein demonstrate that controlling for specific environmental confounders yields a statistically significant improvement in the structural integrity of the synthesized views. This investigation not only provides a robust theoretical foundation for understanding neural rendering artifacts but also offers actionable insights for optimizing data acquisition protocols in the context of digital heritage preservation.