Using Neural Radiance Fields and Gaussian Splatting for 3D reconstruction of aircraft inspections

Abstract: The rapid evolution of machine learning techniques has revolutionized computer vision, particularly with the introduction of Neural Radiance Fields (NeRF) and the optimization of 3D Gaussians for rendering novel scene views. These methods, such as NeRF and Gaussian Splatting, have demonstrated success in synthetic data scenarios with consistent lighting and well-captured scenes. This research explores the feasibility of applying these techniques to images captured by drones conducting aircraft inspections, aiming to automate and optimize the aviation industry. Motivated by the need for more efficient inspection processes, various models were examined and parameters adjusted to assess the performance of NeRFs and Gaussian Splatting in real-world scenarios. Despite the visual shortcomings observed in both NeRF and Gaussian Splatting, Gaussian Splatting came out as more promising for inspection images, outperforming NeRF visually. Quantitative results are presented, using metrics such as SSIM, PSNR and LPIPS to provide a more concrete understanding of the visual performance of the methods. While Gaussian Splatting exhibits promise, it is essential to acknowledge that the current quality of the output falls short of production standards. Looking ahead, with both methods being relatively new, substantial improvements in performance and broader applications are anticipated in the near future.