| This paper investigates the influence of various settings on the accuracy of 3D reconstructions created using Instant-NGP. Using an owl statue as a case study, camera parameters were estimated in COLMAP with different matchers and camera models, and subsequent reconstructions were generated. The accuracy of the resulting models was evaluated by comparing point clouds with a reference photogrammetric model using mean deviation and RMS error. The results showed that a smaller reconstruction volume and an appropriate combination of parameters lead to more accurate and stable results. Certain matcher types, such as the transitive matcher, proved to be unreliable. The proposed methodology enables objective comparison of results and selection of optimal settings for Instant-NGP. In addition to quantitative evaluation, the study also highlights practical aspects of neural reconstruction workflows. The findings contribute to a better understanding of how parameter selection influences not only visual quality but also measurable geometric fidelity. Study also investigates how individual parameters interact to affect both global and local accuracy. Visual inspections complement the numerical analysis to identify regions prone to larger deviations, particularly edges and fine details of the reconstructed object. By combining objective metrics with practical workflow considerations, the methodology provides actionable guidance for researchers and practitioners seeking to optimize Instant-NGP reconstructions. Overall, this work emphasizes the importance of parameter tuning in neural 3D reconstruction and demonstrates a reproducible approach for evaluating geometric fidelity across different configurations. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.