README.md

---
language:
  - en
tags:
  - 3d
  - text-retrieval
  - asset-metadata
  - sqlite
  - faiss
---

# ObjectAtlas Data

A prebuilt asset index and vector library for searching individual 3D objects with natural language. Descriptions from Cap3D, MARVEL-40M+, and TRELLIS-500K are grouped by asset, preserving their annotation sources and original fields.

This bundle contains metadata and description embeddings. Model files can be obtained through the download locations stored in the asset records.

Code and usage examples: [ObjectAtlas on GitHub](https://github.com/unisis-lab/ObjectAtlas).

## Contents

| File | Purpose | Size |
|---|---|---:|
| `data/index.sqlite` | Asset metadata, captions, and vector ID mappings | 13.17 GiB |
| `data/vectors/index.faiss` | FAISS index for description retrieval | 1.43 GiB |
| `data/vectors/vectors.f32` | Full precision description vectors for reranking | 39.46 GiB |
| `extra/extra.sqlite` | Extra asset metadata, archive manifest, and vector ID mappings | 0.14 GiB |
| `extra/vectors/vectors.f32` | Precomputed description vectors for extra assets | 0.29 GiB |

The five files total approximately **54.49 GiB**. Keep each collection's metadata and vector files together to preserve their ID mappings.

| Collection | Assets | Description vectors |
|---|---:|---:|
| Main database | 1,842,649 | 13,792,908 |
| Extra database | 10,840 | 102,896 |

```text
artifacts/
  README.md
  data/
    index.sqlite
    vectors/
      index.faiss
      vectors.f32
  extra/
    extra.sqlite
    vectors/
      vectors.f32
```

## Asset metadata

The `assets` tables in `data/index.sqlite` and `extra/extra.sqlite` share five fields:

| Field | Description |
|---|---|
| `uuid` | Stable local asset identifier |
| `source_dataset` | Original asset dataset, such as `objaverse`, `objaverse_xl`, or `abo` |
| `original_id` | Original model ID, stored as a string with leading zeros and case preserved |
| `captions` | Descriptions grouped under `cap3d`, `marvel_40m_plus`, and `trellis_500k`; each item contains `text` and the original `field` |
| `download_urls` | Available asset file locations; an empty list for extra assets awaiting local installation |

`captions` and `download_urls` are stored as JSON text in SQLite. Missing caption sources use empty lists.

Assets are matched by their confirmed original dataset and model identity. Similar captions, names, or categories do not establish asset identity. A single asset can have multiple description vectors.

## Vector format

| Setting | Value |
|---|---|
| Encoder | `sentence-transformers/all-mpnet-base-v2` |
| Model revision | `e8c3b32edf5434bc2275fc9bab85f82640a19130` |
| Dimensions | `768` |
| Maximum sequence length | `384` tokens |
| Normalization | Unit length |
| Raw format | Headerless, row-major, little-endian float32 (`<f4`) |
| Similarity | Cosine similarity |

The SQLite `captions` table links vector IDs to assets and their descriptions. The `vector_segments` table records each raw vector segment's path, first ID, and row count. The main FAISS index uses the same caption vector IDs.

## Extra collection

The extra collection contains Toys4k, OmniObject3D, and selected ShapeNet categories whose models are distributed in archives. It includes original asset metadata, official archive manifests, and precomputed description vectors. Model archives and extracted model files are not included in this bundle.

## Sources and licenses

Descriptions come from [Cap3D](https://huggingface.co/datasets/tiange/Cap3D), [MARVEL-40M+](https://huggingface.co/datasets/sankalpsinha77/MARVEL-40M), and [TRELLIS-500K](https://huggingface.co/datasets/JeffreyXiang/TRELLIS-500K).

| Original asset source | Official resource |
|---|---|
| Objaverse / Objaverse-XL | [Objaverse](https://objaverse.allenai.org/), [Objaverse-XL metadata](https://huggingface.co/datasets/allenai/objaverse-xl) |
| ABO | [Amazon Berkeley Objects](https://amazon-berkeley-objects.s3.amazonaws.com/index.html) |
| GSO | [Google Scanned Objects](https://research.google/blog/scanned-objects-by-google-research-a-dataset-of-3d-scanned-common-household-items/) |
| HSSD | [HSSD models](https://huggingface.co/datasets/hssd/hssd-models) |
| ShapeNet | [ShapeNetCore v2](https://huggingface.co/datasets/ShapeNet/ShapeNetCore), [ShapeNetCore GLB](https://huggingface.co/datasets/ShapeNet/shapenetcore-glb) |
| Toys4k (extra) | [Official project](https://github.com/rehg-lab/lowshot-shapebias/tree/main/toys4k) |
| OmniObject3D (extra) | [Official project](https://github.com/omniobject3d/OmniObject3D), [OpenXLab](https://openxlab.org.cn/datasets/omniobject3d/OmniObject3D-New) |

The MIT license for ObjectAtlas runtime code does not replace the licenses of the underlying data. Metadata, descriptions, models, and vectors remain subject to the applicable source licenses and attribution requirements. Cap3D annotations are listed under ODC-By, MARVEL under CC BY-NC-SA 4.0, and TRELLIS metadata under MIT; original models remain subject to their publishers' and authors' terms.

Review source licenses before using or redistributing this bundle, including any noncommercial restrictions and attribution requirements. Users are responsible for obtaining access to restricted sources such as ShapeNet and the official Toys4k downloads.