Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Large Scale 3D Gaussian Splatting SLAM
This project aims to create large scale 3D Gaussian Splatting scenes using online robotic data.
Keywords: SLAM, Radiance Fields, 3D Gaussian Splatting, Robotics, Computer Vision
Recent developments in Gaussian Splatting research have enabled their use in real-time SLAM systems, allowing the rapid generation of high fidelity environmental reconstructions. While these methods are fast enough to run on robotic hardware, they are often memory bound; limiting the ultimate size of a scene. Meanwhile offline works have added the capability to divide a scene into subsections for large scale compatibility, capable of accurately capturing city-sized scenes. This project seeks to merge these advances to create a real-time SLAM system that utilizes sub mapping for large scale scene capture. This would involve investigating methods of sub mapping from both SLAM and Radiance Field backgrounds, building off an existing framework to enable more efficient 3DGS SLAM, and testing on real and simulated datasets.
Recent developments in Gaussian Splatting research have enabled their use in real-time SLAM systems, allowing the rapid generation of high fidelity environmental reconstructions. While these methods are fast enough to run on robotic hardware, they are often memory bound; limiting the ultimate size of a scene. Meanwhile offline works have added the capability to divide a scene into subsections for large scale compatibility, capable of accurately capturing city-sized scenes. This project seeks to merge these advances to create a real-time SLAM system that utilizes sub mapping for large scale scene capture. This would involve investigating methods of sub mapping from both SLAM and Radiance Field backgrounds, building off an existing framework to enable more efficient 3DGS SLAM, and testing on real and simulated datasets.
- Literature research and benchmarking existing Visual-SLAM and 3DGS methods.
- Developing the framework, testing on synthetic data, and real world testing.
- Conducting thorough ablation studies on different aspects of the developed method.
- Literature research and benchmarking existing Visual-SLAM and 3DGS methods. - Developing the framework, testing on synthetic data, and real world testing. - Conducting thorough ablation studies on different aspects of the developed method.
- Excellent knowledge of Computer Vision and Python.
- Basic knowledge of machine learning architectures.
- Highly motivated and research-oriented.
- Optional: Experience with Radiance Fields or SLAM systems.
-
- Excellent knowledge of Computer Vision and Python. - Basic knowledge of machine learning architectures. - Highly motivated and research-oriented. - Optional: Experience with Radiance Fields or SLAM systems. -
- Maximum Wilder-Smith, mwilder@ethz.ch
- Vaishakh Patil, patilv@ethz.ch
- Michael Niemeyer, Google Zurich
- Michael Oechsle, Google Zurich
- Keisuke Tateno, Google Zurich
- Maximum Wilder-Smith, mwilder@ethz.ch - Vaishakh Patil, patilv@ethz.ch - Michael Niemeyer, Google Zurich - Michael Oechsle, Google Zurich - Keisuke Tateno, Google Zurich
Large Scale 3DGS SLAM.pdf
