hugging-3d/google-scanned-objects

Google Scanned Objects (WebDataset)

A load_dataset-ready WebDataset packaging of the Google Scanned Objects dataset: high-quality 3D scans of common household items. Each object provides the original mesh/material/texture, five render thumbnails, structured metadata, and a normalized GLB produced with trimesh for direct use in 3D pipelines (e.g. TRELLIS-2 fine-tuning).

• Objects: 1030
• Shards: 43 (data/gso-train-*.tar)
• License: Creative Commons Attribution 4.0 International (cc-by-4.0)

Usage

from datasets import load_dataset

ds = load_dataset("suvadityamuk/google-scanned-objects", split="train")
sample = ds[0]
print(sample["json"]["name"], sample["json"]["category"])
sample["texture.png"]        # PIL.Image (diffuse texture)
sample["thumbnail_0.jpg"]    # PIL.Image (render)
glb_bytes = sample["glb"]    # bytes of a self-contained, normalized .glb

Streaming works too (no full download):

ds = load_dataset("suvadityamuk/google-scanned-objects", split="train", streaming=True)
sample = next(iter(ds))

GLB is a recognized mesh extension in the datasets WebDataset builder, so glb is loaded as a Mesh feature and rendered in the dataset viewer. A recent datasets version is required for the Mesh feature.

Per-sample fields

field	type	description
obj	bytes	original model.obj mesh
mtl	bytes	original model.mtl material
texture.png	Image	diffuse texture
glb	Mesh	normalized, self-contained GLB (see below)
thumbnail_0.jpg .. thumbnail_4.jpg	Image	5 render thumbnails
json	dict	metadata + glb_processing stats

The json field contains: object_id, name, description, version, category, annotations (e.g. brand, gtin, sku), author, license, license_id, license_url, copyright, source, num_thumbnails_original, and glb_processing.

GLB normalization (trimesh)

Produced deterministically per object with trimesh==4.12.2:

1. Copy materials/textures/texture.png next to the OBJ so the MTL's map_Kd texture.png resolves and the diffuse map binds.
2. trimesh.load(obj, process=False, force="mesh") — preserves UVs, flattens to a single mesh.
3. Translate the axis-aligned bounding-box center to the origin.
4. Uniformly scale so the longest AABB edge is 1.0 (fits a unit cube).
5. Preserve UVs (no vertex merge); recompute normals only if missing.
6. Export a self-contained .glb with the texture embedded.

Every object's exact transform (applied_translation, applied_scale), mesh stats (vertices, faces, is_watertight, surface_area_normalized, orig_extents, final_extents, texture_size, glb_size_bytes) and any warnings are recorded in json["glb_processing"] and aggregated in stats/processing_log.jsonl.

Excluded objects

This dataset contains only the 1,030 objects authored by Google (Fuel owner GoogleResearch) in the "Scanned Objects by Google Research" collection. That Fuel collection's listing also includes 16 models owned by other authors; they are intentionally excluded because they are not Google scans, use a different file structure (.dae/.obj, optional textures, 2-6 thumbnails), and carry different licenses (CC-BY-4.0 and CC0-1.0):

• OpenRobotics (13): Bed, Cardboard box, Door handle, Fridge, Hinged door, KitchenCountertop, LampAndStand, Plastic Cup, Shower, SquareShelf, Standard Toilet, Vanity, WhiteCabinet
• LeDSantos (3): ada_lovelace_poster, garage_door, toilet_paper

These appear as status: "failed" rows in stats/processing_log.jsonl (the raw downloader requested them under the wrong owner, yielding invalid archives).

License & attribution

This dataset is a repackaging of the Google Scanned Objects dataset by Google LLC, licensed under Creative Commons Attribution 4.0 International (cc-by-4.0).

"Google Scanned Objects", Copyright 2020 Google LLC, licensed under CC BY 4.0.

You must give appropriate credit when using this data. See the LICENSE file and each object's json["copyright"] / json["license"] fields.