One short walk with a 360 camera. Watch the park build up in 3D around the camera, in the order the camera first saw it, then step into the finished Gaussian splat and look around.
Capture record
How we built it
We keep the sharpest moment in every third of a second from both of the X6's fisheye lenses, used raw with no stitching. COLMAP solves the two lenses together as one rig, so both always face opposite ways from the same spot, and works out where the camera was at every moment.
An open-vocabulary detector and SAM 2.1 mask every person in every frame, the person holding the camera included, plus the pram and hands. LichtFeld Studio fits 1.5 million Gaussians to the rest. Any faint trace of a person left behind is turned grey.
Every part of the splat is tagged with the first frame whose camera saw it close enough. Playing those tags in order replays the park appearing as the walk goes on.
At the end of the walk the points give way to the full Gaussian splat, the same scene drawn photoreal, and you can look around it.
Limits