Revisão sistemática de técnicas de renderização de nuvens de pontos: NeRF, splatting e métodos híbridos

Abstract

One of the main techniques for visualizing real-world objects in virtual environments requires transforming them into point clouds, which makes rendering a computational challenge and has motivated the emergence of multiple approaches in recent literature. In this context, this study conducts a systematic review of point cloud rendering techniques from 2020 to 2025, with the objective of identifying, categorizing, and synthesizing the main approaches and challenges. The methodology followed a rigorous systematic review protocol, with a search on Google Scholar that resulted in 746 initially identified articles and, after three screening stages (title/abstract, eligibility, and full reading), 53 articles selected for in-depth analysis. These studies were organized into three thematic tracks: NeRF Track, focused on methods based on Neural Radiance Fields; Splatting Track, dedicated to classical and modern splatting techniques, with emphasis on 3D Gaussian Splatting; and Hybrid Track, which combines elements from both approaches. The results address three research questions regarding: (RQ1) challenges and visual limitations of the techniques; (RQ2) technological evolution and advances in photorealism; and (RQ3) metrics and evaluation methods employed. In summary, the main challenges concentrate on computational efficiency, visual quality, scalability for large scenes, and precise geometric reconstruction, while evaluation has migrated from error-based metrics (PSNR) alone to richer sets that include structural, perceptual, geometric, and performance measures. The evidence indicates that there is no single dominant representation, but rather different combinations of techniques that consciously balance the trade-offs between visual quality, computational cost, and robustness, contributing to guiding the development of new solutions in point cloud rendering.