Content area
Abstract
Generating three-dimensional (3D) content is pivotal for a wide range of applications across various domains, including entertainment, cultural preservation, virtual reality or augmented reality education environments, as well as medical training spaces. Traditional methods for 3D content generation are time consuming, require expertise and can be expensive. The recent emergence of Radiance Field methods in computer graphics has revolutionized the 3D content generation pipeline. Among these, 3D Gaussian Splatting (3DGS) has been considered to be an effective tool for high-fidelity scene reconstruction. This study evaluates the effectiveness of the 3DGS technique for generating realistic 3D content and examines various data acquisition techniques that enhance its robustness. Furthermore, we propose a novel sharpness-guided, overlap-constrained frame selection algorithm that improves both reconstruction quality and computational efficiency. Our framework demonstrates significant performance gains for the computation required with 3DGS, particularly with large-scale datasets—reducing the COLMAP preprocessing time by up to 32.2× and improving reconstruction quality, with PSNR scores increasing by up to 13.86%. In general, radiance field techniques, like 3DGS, are approachable content generation techniques capable of recreating specific spaces. Computational improvements and informed acquisition techniques can help guide that progress.
