GaussianNexus: Room-Scale Real-Time AR/VR Telepresence with Gaussian Splatting
summary
System builds a photo‑real, fully navigable room using Gaussian Splatting from a 3‑minute scan, sampling 300–400 images, 30 min training and per‑object Saga models that take ~45 min plus ~15 min for manual segmentation. Live 2D updates project a top‑mounted 360° camera onto selected surfaces; live 3D updates avoid retraining by tracking segmented objects with YOLOv11, 4‑DOF tracking, 2.3 FPS, 0.95 s latency, aligning scenes with QR anchors. Evaluations report <10 cm positional and <6° rotation error and user study shows high presence.
AI-written summary · about
chapters (12)
- 0:00Introduction and speakers
- 0:45What is telepresence
- 2:56Gaussian Nexus overview
- 3:40Live demo and features
- 4:21Scene preparation process
- 5:23Spatial alignment with QR codes
- 5:45Handling 2D scene updates
- 6:27Enabling 3D object updates
- 8:51System performance and latency
- 9:32User study results
- 10:14Future improvements
- 10:54Q&A session
related talks
41mComputer Vision for Remote AR
20mInteractive Room Capture on 3D-Aware Mobile Devices
39mDesigning for the Living Room: The Spatial Architecture of Mixed Reality
16mInteractive 360-Degree Glasses-Free Tabletop 3D Display
21mAnnexing Reality: Enabling Opportunistic Use of Everyday Objects as Tangible Proxies in ...
27mHoloportation: Virtual 3D Teleportation in Real-time
43mVirtual Reality; Past, Present and Future
26mFrom Panoramic Vision to Lightweight 3D Worlds for Semantic Visual Tracking