Análise comparativa entre NeRF, Gaussian Splatting e versões do framework nerfstudio para reconstrução de cenas 3D

Abstract

This work presents an in-depth comparative analysis of 3D reconstruction methods based on Neural Radiance Fields (NeRF) and Gaussian Splatting (GS), focusing on the Nerfacto, Splatfacto, and their variant models. The investigation was conducted using two versions of the Nerfstudio framework: the original release and an enhanced version proposed in this study, which implements a frame-selection strategy optimized by a Laplacian operator to prioritize image sharpness. The main objective was to evaluate the performance of these models in terms of visual quality (PSNR, SSIM, and LPIPS metrics) and computational efficiency (training time, VRAM/RAM usage, and FPS), using a diverse dataset encompassing indoor, outdoor, and human-centered environments. The results confirmed the superiority of Gaussian Splatting–based methods over NeRF-based approaches across virtually all criteria. Models from the Splatfacto family demonstrated not only higher visual quality but also drastically greater rendering efficiency, achieving frame rates over 100 times higher than Nerfacto. The Splatfacto-big model stood out for its maximum visual fidelity, while Splatfacto-w-light emerged as the most balanced solution for real-time applications due to its low computational cost. The comparison between framework versions revealed that the optimized frame-selection strategy, while theoretically advantageous for reducing blur, did not guarantee superiority over the original version’s uniform sampling. This outcome highlights the robustness of the standard implementations for diverse scenarios