Optimizing Multi-Camera Rigs for 3D Gaussian Splatting – A Virtual Simulation Tool

Sammanfattning

3D Gaussian Splatting (3DGS) has emerged as a computationally efficient method for high-quality novel view synthesis and volumetric scene reconstruction. However, the quality of the resulting models depends strongly on the capture setup and camera distribution used during data acquisition. This thesis investigates how the configuration of a cylindrical multi-camera rig influences the reconstruction quality of 3DGS models. To enable large-scale experimentation, a virtual evaluation tool was developed using Three.js, capable of automatically generating synthetic datasets, executing Structure-from-Motion reconstruction through COLMAP, training 3DGS models in LichtFeld Studio, and evaluating the results through image similarity metrics. The tool was used to perform grid-search experiments across hundreds of rig configurations while varying parameters such as camera density, rig radius, vertical spacing, tapering, and sparsity patterns. The findings indicate that while maximizing camera density consistently increases the total number of reconstructed Gaussians, the perceived visual quality and standard image similarity metrics reach a point of diminishing returns relatively early. Notably, deploying a checkerboard sparsity pattern proved to be an optimal trade-off, significantly reducing the required camera count while preserving core reconstruction fidelity. Additionally, the rig's radius was found to be a critical factor for structural robustness; smaller radii limited environmental context, which frequently led to scene duplication errors. The study also highlights the fragility of sparse reconstruction of virtual scenes, especially regarding object symmetry, lighting, and environmental context, all of which caused multiplied point clouds and poor camera position estimates, even when initialized with prior camera poses. Furthermore, there were limitations in the evaluation methodology. Since validation images originated from viewpoints already present in the training data, PSNR and SSIM values were likely inflated, restricting the ability to fully evaluate the effect of angular spacing between cameras. Ultimately, this work demonstrates that highly efficient volumetric capture systems can be designed without strictly maximizing camera count, and provides a reproducible virtual evaluation tool to guide future physical rig development.