Skip to content

splats

splats #

Splats — 3D Gaussian splats from posed captures, as recipes.

A splat is trained on posed images from a seed, and each of those is its own memoized step, so a scene's splat and its sessions' share what they compute:

  • .datasets — posed frames from each capture shape: scene_dataset (a scene capture's keyframes), video_dataset (every stride-th frame of a posed video), panorama_dataset (a sweep's frames placed by GPS and compass, several sweeps at one location sharing a world); one rule turns a decoded frame into a posed image.
  • .seeds — the coloured world points the gaussians start from: depth_seed (the frames' pixels through their captured depth), distance_seed (at a constant distance instead), splat_centers (the centres of trained splats, so a scene's splat starts from its sessions') and merged (several seeds as one).
  • .training — train: the splat, trained by nerfstudio in the GPU interpreter under cgr.gpu_lock(), supervised on each frame's measured depth, its progress a logged count; and the driver it runs there.
  • [.viewer][] — SplatViewer: the page component that orbits a trained splat in the browser, drawn by gsplat; exported at the cgr root.

Camera geometry (Pose as a matrix, pixels as rays) is cgr.sfm.camera, the layer beneath.

This package is the cgr.splats namespace:

frames = cgr.splats.scene_dataset(session.capture)
splat = cgr.splats.train(frames, cgr.splats.depth_seed(frames))