3D Gaussian Splatting

Comparing SLAM frontends for online 3D Gaussian Splatting, comparing how pose estimation affects reconstruction fidelity.

Final reportProject code
  • OpenVINS
  • ORB-SLAM3
  • VGGT-SLAM
  • 3D Gaussian Splatting
  • Python

Gallery

Summary

This project compared how different SLAM estimators affect online 3D Gaussian Splatting reconstruction quality. We used the RPNG AR Table Dataset, and ran OpenVINS, ORB-SLAM, and VGGT-SLAM to estimate point clouds and pose estimates With this visual odometry, we reconstructed the scene using 3D Gaussian splats. We compared the estimators, finding the ORB-SLAM had the highest fidelity for both positon and orientation. My role was to run the OpenVINS estimator, generating one set of point clouds, and matching it to the camera frames.

The reconstructed scene using 3D gaussian splats and ORB-SLAM, which provided the highest fidelity reconstruction.
Training preview using ORB-SLAM for pose estimates.
Side-by-side estimator comparison.
Position error over time for the SLAM estimators.
CDF view of position error, showing how often each estimator stayed within a given error bound.