Downstream Pipelines

Downstream Pipelines

A splatpack is reconstruction-ready input. This guide outlines how to take one into a radiance field reconstruction. The right path depends on which mode you captured in.

Pick your path

Choose the mode you captured in to see the recommended route.

LiDAR → Gaussian splatting

Use the on-device COLMAP model — skip structure-from-motion entirely.

  • Locate COLMAP_Text_Model/ inside your LidarSeries_<date> folder.
  • Point your Gaussian splatting trainer at the model — images under images/, sparse model under sparse/0/.
  • Poses come straight from ARKit: fast and deterministic, no feature matching.

LiDAR → Gaussian splatting (fastest path)

LiDAR captures include an on-device COLMAP model (COLMAP_Text_Model/) — camera poses, images, and a sparse point cloud — so you can skip structure-from-motion entirely.

  1. Locate COLMAP_Text_Model/ inside your LidarSeries_<date> folder.
  2. (Optional) Inspect it in COLMAP: File → Import model, with the image path set to COLMAP_Text_Model/images/.
  3. Point your Gaussian splatting trainer at the model — images under images/, the sparse model under sparse/0/.

Because the poses come straight from ARKit, this path is fast and deterministic — no feature matching, no SfM solve.

Photo or Video → COLMAP → Gaussian splatting

Photo and video captures don’t carry camera poses, so you reconstruct them the standard way: recover poses with COLMAP, then train. For video, extract frames from wide.mov and ultra.mov first (using both lenses gives broader coverage); for photo, use the wide_* and ultra_* images directly.

  1. Feature extraction
  2. Feature matching
  3. Mapping (structure-from-motion) → a sparse model
  4. Take the sparse model + images into your Gaussian splatting trainer.

COLMAP’s Automatic Reconstruction (GUI) or colmap automatic_reconstructor (CLI) runs steps 1–3 in one pass.

LiDAR point cloud → CloudCompare

To inspect or use the fused LiDAR point cloud directly, open lidar_pointcloud_world_xyz.ply in CloudCompare. See the CloudCompare guide for details.

Notes

Use both lensesIn photo and video, the 0.5× ultra-wide and 1× wide streams cover the scene from two synchronized viewpoints — feeding both into one reconstruction improves coverage and parallax.
Filter weak framesEach frame’s per-frame metadata carries a quality assessment; dropping low-quality frames before reconstruction can improve results.
Coordinate systemLiDAR outputs (the COLMAP model and the point cloud) are in the ARKit world frame: right-handed, gravity-aligned, +Y up, units in meters.