| Project Domains | Mentors |
|---|---|
| Computer Vision, Machine Learning, 3D Reconstruction, Gaussian Splatting | Sarayu Anantharaman, Shashvat Prabhu |
Project Description
This project’s goal is taking a regular video walking through VJTI’s campus and turning it into a photorealistic, navigable 3D environment where you can walk around like a First-Person game.
What makes this project interesting is that you learn to tackle one of the hardest problems in 3D Reconstruction: efficiently solving camera position tracking of every frame over long video trajectories without drift. This is referred to as “Long Context 3d Scene Reconstruction”.
The camera positions and depth maps across the trajectory supervise the training of a 3D Gaussian Splat. We then integrate it into a pose-graph navigation loop for smooth exploration.
This is a direct step into cutting-edge frontier research on 3D Reconstruction and Gaussian Splatting. High-quality interactive 3D scenes have a direct application in perception driven robotics, autonomous driving, embodied AI and digital twins. By the end of the project, you will have a cool demo to showcase, as well as a solid intuition on the problems that are being worked on in this domain.
Resources
3D Gaussian Splatting for Real-Time Radiance Field Rendering
LoGeR Paper