suvadityamuk/amazon-berkeley-objects

Amazon Berkeley Objects

This repository is a viewer-compatible Hugging Face packaging of the official Amazon Berkeley Objects (ABO) release. ABO contains 147,702 multilingual product listings, 398,212 unique catalog images, 586,584 turntable images, 7,953 high-quality glTF 2.0 models, material-benchmark renders and geometry buffers, fine-grained 3D part labels, and the official multi-view retrieval splits.

The data content is the official CC BY 4.0 release. This mirror changes only the packaging: metadata and media are grouped into typed Parquet shards so the Hugging Face Dataset Viewer and datasets streaming APIs can consume them. No claim of endorsement by Amazon or UC Berkeley is made.

Recommended entry point

The default objects config is the product-level index and the easiest way to browse ABO. Each row joins a listing to its 256 px main image, native GLB mesh when available, and flags/foreign keys for spins, part labels, and material benchmark records.

from datasets import load_dataset

objects = load_dataset(
    "suvadityamuk/amazon-berkeley-objects",
    "objects",
    split="train",
    streaming=True,
)
first = next(iter(objects))
print(first["item_id"], first["item_name"], first["has_3d_model"])

Use a modality config for complete training data:

images = load_dataset(
    "suvadityamuk/amazon-berkeley-objects",
    "images_small",
    split="train",
    streaming=True,
)
models = load_dataset(
    "suvadityamuk/amazon-berkeley-objects",
    "models_3d",
    split="train",
    streaming=True,
)

Native 3D preview uses the Hugging Face Mesh() feature and therefore requires a recent datasets release with mesh support.

Configs

Config	Row unit	Main fields
objects	One product	Listing summary, Image() main image, optional Mesh() GLB, modality flags
listings	One product listing	Display fields plus lossless raw_listing_json
images_small	One catalog image	256 px Image(), image id, dimensions, official path
images_original	One catalog image	Original-resolution Image(), image id, dimensions, official path
spins	One turntable frame	Image(), spin id, image id, azimuth
models_3d	One 3D product	Native Mesh() GLB and geometry/material statistics
part_labels	One connected/convex part mesh	Original OBJ bytes, converted Mesh() GLB preview, semantic label and relations
benchmark_material	One model viewpoint	Base color, metallic/roughness, normal, segmentation and three render Image() columns; EXR depth bytes; camera data
mvr	One retrieval example	Official class/group/image ids, source path, metadata and split

The source_metadata/ directory retains the official top-level README, download URL list, compressed metadata tables, benchmark split files, and sample indices.

Format notes

• JPEG and PNG fields use Image() and render in the Dataset Viewer.
• Official GLB files use Mesh() and render in the 3D viewer.
• Part-label OBJ files are preserved byte-for-byte in source_obj; an additional GLB conversion in mesh enables browser preview.
• OpenEXR depth maps are preserved byte-for-byte in depth_exr. EXR does not have a native Hub image preview, so each row also exposes its corresponding JPEG/PNG render and material maps.
• Empty strings and -1 labels indicate fields not supplied by the official source. A non-empty conversion_error means an OBJ was preserved but could not be converted for mesh preview.

Source dataset sizes

The extracted official release used for this mirror occupies approximately 599 GiB:

• listings: 84 MiB
• small catalog images: 3.6 GiB
• original catalog images: 111 GiB
• spin frames: 41 GiB
• 3D models: 155 GiB
• part labels: 11 GiB
• material benchmark: 277 GiB

Typed Parquet configs can be larger in aggregate because objects embeds preview media already present in modality-specific configs and part labels contain both source OBJ and preview GLB representations.

License

The official release is licensed under the Creative Commons Attribution 4.0 International License. The complete license text is included as LICENSE-CC-BY-4.0.txt.

You must give appropriate credit, link to the license, and indicate whether changes were made. You may not imply endorsement by the licensors.

Required attribution

The official archive README requires the following credit:

Credit for the data, including all images and 3D models, must be given to: Amazon.com

Credit for building the dataset, archives and benchmark sets must be given to: Matthieu Guillaumin (Amazon.com), Thomas Dideriksen (Amazon.com), Kenan Deng (Amazon.com), Himanshu Arora (Amazon.com), Jasmine Collins (UC Berkeley), and Jitendra Malik (UC Berkeley).

The current official project page additionally credits Arnab Dhua, Xi (Brian) Zhang, Tomas Yago-Vicente, and Shubham Goel.

Citation

Please cite the official CVPR 2022 paper when using ABO:

@InProceedings{Collins_2022_CVPR,
  author    = {Collins, Jasmine and Goel, Shubham and Deng, Kenan and
               Luthra, Achleshwar and Xu, Leon and Gundogdu, Erhan and
               Zhang, Xi and Yago Vicente, Tomas F. and Dideriksen, Thomas and
               Arora, Himanshu and Guillaumin, Matthieu and Malik, Jitendra},
  title     = {ABO: Dataset and Benchmarks for Real-World 3D Object Understanding},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and
               Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2022},
  pages     = {21126--21136}
}

The fine-grained part labels were released for the 3D Vision and Modeling Challenges in eCommerce, ICCV 2023 workshop challenge.

Official resources

• ABO project and downloads
• CVPR 2022 paper
• CC BY 4.0