3D Gaussian Splatting

We present an end-to-end pipeline for converting multi-view RGB captures into object-centric, physics-enabled virtual reality scenes rendered with 3D Gaussian Splatting. Existing 3DGS reconstructions provide high visual fidelity but are typically scene-level representations without explicit object structure or physical affordances. Our method bridges this gap by combining structure-from-motion calibration, 3DGS reconstruction, prompt-driven instance segmentation, geometry-guided mask propagation, and per-Gaussian object labeling. The resulting representation separates the reconstructed scene into independently manipulable Gaussian assets and a static background. To support physically plausible interaction, the pipeline conservatively completes occlusion-induced missing regions and derives simplified mesh proxies for collision and rigid-body simulation. At runtime, each invisible proxy is synchronized with its corresponding visible Gaussian asset, preserving the appearance quality of neural rendering while enabling object grabbing, movement, and collision response in VR. This framework provides a practical approach for transforming image-based scene captures into interactive immersive environments, narrowing the gap between photorealistic neural reconstruction and real-time physics-based manipulation.