CloudAnalyzer / 3DGS Evaluation

Render a Gaussian Splat. Score photometric + geometry in one gate.

ca rendered-evaluate drives a 3DGS PLY through gsplat at nerfstudio camera poses, then chains PSNR / SSIM / LPIPS with optional Chamfer / AUC / F1 against a reference scan. The bundled synthetic-room demo below is fully reproducible from the repository — no external datasets required.

Command Reference Demo Hub Demo Dataset

Sample renders (synthetic-room)

Reference views generated from gaussians_dense.ply via gsplat.

One-command regression check

pip install "cloudanalyzer[gs]"

ca rendered-evaluate benchmarks/3dgs/synthetic-room/gaussians_dense.ply \
  benchmarks/3dgs/synthetic-room/reference \
  --cameras benchmarks/3dgs/synthetic-room/transforms.json \
  --reference-pointcloud benchmarks/3dgs/synthetic-room/reference.pcd \
  --metrics psnr,ssim,lpips \
  --report rendered-report.html

What you get

PSNR / SSIMPhotometric fidelity per view

LPIPSPerceptual distance (optional [gs])

Chamfer / AUCGeometry QA vs reference scan

HTML reportCombined gate for CI / PR comments

Camera poses: nerfstudio transforms.json or COLMAP text exports.
Renderer: gsplat behind pip install "cloudanalyzer[gs]".
Geometry-only path still available via ca geometry-evaluate.