GSO -- Vtech Roll & Learn Turtle (scanned toy)

gso_household · Google Scanned Objects (GSO) · downloaded from the dataset (4 file(s))

Ground-truth mesh

Downloaded from the dataset. No local renders yet, so there is no viewer and no GT depth or normals — the acquisition plan below describes how to render it.

Why this sample is in the benchmark

Real structured-light scan: real albedo, real scan noise, real asymmetry, and no clean symmetry to hide behind. It is the bridge between the analytic tier and full scene captures -- and it is the one external sample that downloads without any manual step, so the "real data" path can be exercised end to end.

Swap `uid` for any other name from the Fuel collection listing to get a different household object; the URLs are templated from it.

Ground truth

GT type
scanned
licence
CC-BY 4.0
source
Google Scanned Objects (GSO)
builder
- {'uid': 'Vtech_Roll_Learn_Turtle'}
views
40 × 512px (orbit, fov 50°)
lighting
provided

Acquisition plan

kind url, automatic True, estimated size 8 MiB

  1. List the collection's objects (verified REST endpoint) 
    curl -s 'https://fuel.gazebosim.org/1.0/GoogleResearch/collections/Scanned%20Objects%20by%20Google%20Research/models?per_page=1000'
  2. Inspect one object's file tree 
    curl -s 'https://fuel.gazebosim.org/1.0/GoogleResearch/models/Vtech_Roll_Learn_Turtle/tip/files'
  3. Download the ground-truth mesh (and friends) directly 
    curl -L --create-dirs -o gso_household/meshes/model.obj 'https://fuel.gazebosim.org/1.0/GoogleResearch/models/Vtech_Roll_Learn_Turtle/tip/files/meshes/model.obj'
  4. Mirror with one archive per object on Hugging Face
    https://huggingface.co/datasets/suvadityamuk/google-scanned-objects
  5. Dataset announcement and citation
    https://research.google/blog/scanned-objects-by-google-research-a-dataset-of-3d-scanned-common-household-items/

Google Scanned Objects: 1030 structured-light scans of household items, CC-BY 4.0, ~13 GB for the whole collection. Each object ships as a textured mesh -- real sensor noise, real albedo, single capture lighting. The mesh is the ground truth you want; to vary illumination you must relight it yourself.

Comparisons

No reconstruction compared yet. Run geombench compare gso_household --mock or point it at a real 3DGS output.