idirectships/splat-cupr-001

Cuprite on Copper Specimen (cupr-001) — Gaussian Splat

Cuprite on native copper specimen from Ray Mine, AZ. Reconstructed using Gaussian Splatting after two capture sessions — the first (car seat background) failed COLMAP registration (4/25 frames); the second (dark matte background, 123 frames) achieved 115/123 registration.

Specimen Details

• Mineral: Cuprite on native copper
• Location: Ray Mine, AZ
• Listed on: Borussia Minerals

Capture Sessions

Session	Frames	Background	COLMAP Registration	Result
v1	25	Car seat	4/25 (16%)	Unusable
v2	123 (IMG_1934–2057)	Dark matte	115/123 (94%)	Primary dataset

Directory Structure

images_nobg/      — background-removed images from v2 session
v3-capture/       — third capture session
  runs/
    v3-mama-15k/
      output.ply  — 15,000 iteration Gaussian Splat from v3 capture
runs/
  prior-attempts/
    cuprite-copper-4view.ply      — 4-view only (v1 session, 45MB) — reference only
    cuprite-v2-colmap-7k.ply      — v2 COLMAP 7k (41MB)
    cuprite-v2-dust3r-60-7k.ply   — v2 DUSt3R 60-frame subset 7k (47MB)
    cuprite-v2-local-15k.ply      — v2 local 15k
  v1-mama-7k/
    output.ply    — 7,000 iteration Gaussian Splat
  v1-mama-15k/
    output.ply    — 15,000 iteration Gaussian Splat
  v1-mama-30k/
    output.ply    — 30,000 iteration Gaussian Splat (438MB, best run)
  2026-03-20-15000-MamaBerr/
    output.ply    — Mama 15k run
  v1-mama-15k.ply — standalone 15k PLY (49.7MB)
image-report.json
notes.md

PLY Iterations

Run	Iterations	Method	Node	Size	Notes
prior-attempts/cuprite-copper-4view.ply	7k	COLMAP	GUS	45MB	v1 session, 4 views only
prior-attempts/cuprite-v2-colmap-7k.ply	7k	COLMAP	GUS	41MB	v2 session, jagged/floaters
prior-attempts/cuprite-v2-dust3r-60-7k.ply	7k	DUSt3R	GUS	47MB	60-frame subset comparison
v1-mama-7k/output.ply	7k	COLMAP	Mama	—	Standard run
v1-mama-15k/output.ply	15k	COLMAP	Mama	~50MB	Good quality
v1-mama-15k.ply	15k	COLMAP	Mama	49.7MB	Standalone copy
v1-mama-30k/output.ply	30k	COLMAP	Mama	438MB	Best run
v3-capture/runs/v3-mama-15k/output.ply	15k	COLMAP	Mama	—	v3 capture session

Training Pipeline

1. Structure from Motion: COLMAP (v2: 115/123 frames registered) and DUSt3R (for comparison)
2. Background removal: Applied before training
3. Gaussian Splat training: opensplat on Mama (Mac Mini M4)

Known Issues

• Specular reflections on copper matrix confuse SfM — diffuse lighting recommended for reshoot
• Top face has weak coverage — targeted reshoot of top angle would improve reconstruction
• Car seat background in v1 caused catastrophic COLMAP failure (always use dark matte background)

Viewing PLY Files

Use SuperSplat to view .ply Gaussian Splat files in-browser.