Spacetime Gaussians (CVPR 2024) can render moving people from any viewpoint in real time, but only if you feed it footage from a calibrated research rig. I wanted to see what happens with 24 cameras and no calibration: a person, a ball, a room. This project is the pipeline that gets my own videos into the method: COLMAP for camera poses, a converter into the format the loader expects and a GPU box that took more setup than the training itself. The result moves and you can fly the camera round it. It is also blurry and I now know at least one reason why.
