3D Gaussian Splatting · End-to-End Scene Reconstruction

Bamburgh Castle Reconstruction with gsplat

Colab Notebook · ~55 min
Google Colab Notebook
Bamburgh Castle Reconstruction with gsplat
Python · ~55 min
Open in Colab
Lab Objectives
1
Explain why gsplat is a viable commercial alternative to the original gaussian-splatting repository, citing the specific license clauses that create the restriction.
2
Run the full YouTube → COLMAP/GLOMAP → gsplat pipeline on the Bamburgh Castle fly-through and confirm PSNR within 1 dB of the Lab 6 baseline.
3
Interpret gsplat's built-in per-checkpoint PSNR/SSIM/LPIPS output and compare it against the Lab 6 reference values.
4
Verify that gsplat outputs a 59-property PLY schema identical to the original gaussian-splatting, and convert it to a .splat file using the same byte-packing procedure.
5
Load the exported .splat file in a WebGL viewer and identify any visual differences from the Lab 6 reconstruction.
6
Answer the §12 discussion questions connecting the license distinction to real commercial deployment decisions.

Overview

This lab runs the identical Bamburgh Castle reconstruction pipeline as Lab 6 — same video source, same COLMAP/GLOMAP SfM — but replaces the original gaussian-splatting training step with gsplat, a community reimplementation under the MIT license.

The goal is two-fold: demonstrate that the license distinction does not affect reconstruction quality, and gain hands-on familiarity with gsplat's built-in evaluation output (PSNR/SSIM/LPIPS at configurable checkpoints), which the original codebase does not provide.

Prerequisites

Complete Lab 6 before running this lab. You will be comparing outputs directly against the Lab 6 baseline.

What Changes From Lab 6

The pipeline is identical up to and including GLOMAP reconstruction:

Video download → Frame extraction → COLMAP feature extraction
→ COLMAP exhaustive matching → GLOMAP reconstruction

At the training step, gaussian-splatting/train.py is replaced with:

python gsplat_repo/examples/simple_trainer.py default \
  --data_dir /content/data \
  --data_factor 1 \
  --result_dir /content/results/bamburgh \
  --max_steps 30000 \
  --eval_steps 7000 15000 30000 \
  --save_steps 30000 \
  --disable_viewer

gsplat reads the same COLMAP-format sparse/0/ output and writes a .ply with the identical 59-property schema.

Evaluation

Unlike the original codebase, gsplat logs PSNR, SSIM, and LPIPS to a JSON file at each --eval_steps checkpoint. §7 of the notebook parses and plots these metrics.

Reference values (Lab 6 baseline, no depth priors):

  • PSNR: ~29.6 dB
  • SSIM: ~0.88
  • LPIPS: ~0.18

Discussion

Section §12 of the notebook poses four discussion questions connecting the hands-on pipeline to the commercial licensing considerations covered in the Module 1 license reading. Work through these after completing the reconstruction.