Splatting a cell tower
Taking 3D Gaussian Splatting from a room to a full BTS site, and what kept rendering on budget
Most Gaussian Splatting demos live in a room: a desk, a plant, a bicycle. A BTS site is a different animal: a tall, thin structure, a lot of sky, and cameras that never quite agree on exposure.
The problem
The brief: a digital twin operators could fly through, built from a fixed set of drone and ground images. Render speed mattered as much as fidelity: a twin that takes a minute per frame is a slideshow.
We started from a standard pipeline: structure-from-motion for poses, then optimisation of a few million Gaussians.
Thin structures and empty sky are where splatting quietly gives up
Where it broke
Three failure modes showed up in the first hour: floaters in the sky, antennas dissolving into fog, and memory that grew faster than quality did.
Three tricks
- Sky masking before optimisation, so the model stops inventing geometry in empty air
- Scale-aware densification for thin parts like antennas and cables
- Aggressive pruning on a schedule, with a hard memory budget
What's next
We’re turning the experiments into a short paper. If you work on large-scene reconstruction, let’s compare notes.
