Capturing indoor scenes
Indoor laps through Levine and Moore pair color frames with aligned depth images, camera intrinsics, timestamps, and robot poses. These recordings provide appearance and geometry inputs for scene reconstruction.
3D vision · Robotics
F1-3DGS
RGB-D scene reconstruction
Reconstructing indoor environments with 3D Gaussian Splatting from RGB-D lap recordings, combining calibrated images, depth, and robot poses with dataset filtering and reconstruction inspection.
Preparing reliable inputs
Associate captured frames with ROS poses in the map frame. Filtered datasets retain valid image, depth, and pose samples and record skipped transform failures; trajectory plots help inspect spatial coverage across repeated laps.
Gaussian reconstruction
Saved Gaussian models at 7,000 and 30,000 training iterations preserve position, scale, rotation, opacity, and appearance parameters. These outputs capture the reconstructed scene as a collection of 3D Gaussians.
Inspecting the scene
An Open3D viewer loads reconstruction outputs for geometric inspection. Recorded trajectories and RGB inputs provide context for reviewing the captured environment and reconstruction.
Project gallery
RGB input
RGB input from a Levine lap capture.
Visualization coming soon
Gaussian reconstruction
Rendered scene and reconstruction walkthrough — coming soon.