From Snapshots to VR: Evaluating 3D Gaussian Splatting for Accessible 3D Reconstruction

Sammanfattning

3D Gaussian Splatting (3DGS) has emerged as the latest breakthrough in 3D reconstruction since it can provide photorealistic 3D reconstruction with real-time rendering. Despite its increasing popularity, it lacks an accessible and clear end-to-end workflow, and the use for it in interactive and immersive applications remains largely unexplored. Furthermore, with the constant development of new 3DGS-based methods, it becomes unclear which is the best approach. To address these issues, this Thesis first compares three key 3DGS methods — Vanilla 3DGS, 3DGS MCMC, and 2DGS — and ncludes a novel evaluation of their performance from unseen viewpoints. Then, it delivers a ready-to-use pipeline from smartphone scanning to VR interaction, along with supporting manuals and scripts to improve usability and accessibility for non-expert users. Additionally, it describes an intuitive interaction concept for two proof-of-concept VR applications that showcase the potential of 3DGS objects and environments in VR. Finally, it identifies and analyzes the key limitations of 3DGS and proposes strategies to address them. All in all, this Thesis attempts to bridge the gap between the theoretical foundation of 3DGS, practical information about implementations, and the use of such techniques for VR interaction.