3D Scene Understanding

3D Scene Understanding

Arun Reddy Anugu, Kashish Gandhi, Manan Shah, Srinath Ravi

Subsections

• 2D VLMs for 3D VQA
• 4D reconstruction From Monocular Video
• RefineGS: Generalizing 3D Gaussian Splatting for Out of Distribution Views

Motivation

Humans infer rich 3D structure from limited 2D observations, but enabling this in machines remains challenging. While most data is 2D, tasks like reasoning, interaction, and simulation require accurate and dynamic 3D representations. We bridge this gap by

• Enabling 3D question answering with 2D vision-language models
• Reconstructing dynamic 4D scenes from monocular videos
• Improving real-time 3D representations via Gaussian splatting for clean, high-FPS rendering