zeyuanyin/Dome-Objaverse

<div align="center">

Dome-Objaverse

Multi-view renders of 83,296 Objaverse objects — 48 views each, on a camera dome of 4 elevations &times; 12 azimuths.

<p align="center"> <a href="https://huggingface.co/datasets/zeyuanyin/Dome-Objaverse"><img src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-FFD21E?style=for-the-badge&logoColor=black" alt="Hugging Face Dataset"></a> &nbsp;&nbsp; <a href="https://zeyuanyin.github.io/Dome-Objaverse/"><img src="https://img.shields.io/badge/%F0%9F%8C%90%20Viewer%20Website-Online%20Demo-2f6df6?style=for-the-badge" alt="Viewer Website"></a> &nbsp;&nbsp; <a href="https://github.com/zeyuanyin/Dome-Objaverse"><img src="https://img.shields.io/badge/GitHub-Code-181717?style=for-the-badge&logo=github" alt="GitHub Code"></a> </p>


The 48 views in order: three azimuth rings at elevations 0°, 30° and 60°, plus a top-down ring at 90°. The highlighted camera on the dome is the one that took the image on the left.

</div>

Rendering is the computationally demanding bottleneck of multi-view 3D datasets — typically over 100,000 CPU core hours for 83k objects. Our rendered pixels already exist: 512×512 RGBA plus per-view normal and depth maps, ~1.5 TB for the primary split and freely downloadable, each object paired with its Objaverse UID and a Cap3D caption. Dome is literal for the primary dome_objaverse split: all cameras sit on the upper hemisphere (elevations 0° to 90°). The accompanying gobjaverse_parquet split preserves GObjaverse's original trajectory, which dips to elevation −90° (see "Splits & Metadata" below).

• —License: CC-BY-4.0, ~1.5 TB for the primary split, parquet-packaged.
• —Viewer: zeyuanyin.github.io/Dome-Objaverse — search all 83,296 captions, browse every view, and see interactive 3D fused point clouds in-browser before downloading anything.
• —Code: loading/decoding scripts, the rendering pipeline, and full schema docs live at github.com/zeyuanyin/Dome-Objaverse.

Comparison with other Objaverse render sets

Two things distinguish Dome-Objaverse: a regular, fully-specified camera grid (four rings of identical 12-step azimuth sweeps, every view addressable by view_id), and the fact that the pixels are actually published rather than left to the user to render.

Splits & Metadata

Both splits cover the exact same 83,296 objects with zero missing objects.

• —`dome_objaverse`: view_id 0–11 / 12–23 / 24–35 / 36–47 are elevation 0° / 30° / 60° / 90° respectively, each a 12-step azimuth sweep 0°, 30°, ..., 330°.
• —`gobjaverse_parquet`: camera parameters are read dynamically per-view from the embedded 00000.json ... 00039.json metadata files.

metadata/objects.parquet (5 MB, ships in this repo) is the join table, 83,296 rows with 100% coverage. The columns you'll actually use: object_id (e.g. "0/10228"), objaverse_uid, glb_path, caption, camera_distance. It also carries group_id/index_id (the two halves of object_id) and sheet_id/sheet_pos (internal thumbnail-sheet coordinates used only by the web viewer) — safe to ignore for training. Legacy per-field JSON files (camera_distances.json, gobjaverse_index_to_objaverse.json, text_captions_cap3d.json, cobj_done_list.json) are also included; their content is already consolidated into objects.parquet.

Rendering

Full config in pipeline/2-render/blender_script.py; a decode example is in pipeline/4-usage/load_views.py. Two things that will silently produce wrong results if assumed otherwise:

• —⚠️ Camera distance is per object, 1.50–2.00 (see camera_distance in objects.parquet), not a fixed constant.
• —⚠️ `nd_png` (16-bit normal + depth) channel order is reversed: R = normal z, G = normal y, B = normal x, alpha = depth (alpha / 65535 × 5.0, planar along the camera axis). Use cv2.imdecode(..., cv2.IMREAD_UNCHANGED), not PIL — it silently downcasts to 8-bit.

License

Code in this repository is Apache-2.0 (see LICENSE), matching AI2's objaverse-rendering, which render/ derives from. The rendered image data on Hugging Face is CC-BY-4.0. Individual Objaverse source meshes carry their own licenses — use objaverse_uid in metadata/objects.parquet to look them up.

Attribution

• —Source meshes: Objaverse (Allen Institute for AI / AI2).
• —Original camera convention and the 280k render set: GObjaverse / RichDreamer.
• —The curated 83k-object subset comes from ashawkey/objaverse_filter.
• —Captions: Cap3D.
• —Rendering scripts derive from AI2's objaverse-rendering (Apache-2.0; see pipeline/2-render/LICENSE).

Please also cite the original Objaverse / GObjaverse sources and state which split(s) you used.

Citation

This dataset is an extension of our NeurIPS 2025 publication, TRIM: Scalable 3D Gaussian Diffusion Inference with Temporal and Spatial Trimming. If you find it helpful, please consider citing:

@inproceedings{
    Yin2025TRIM,
    title={{TRIM}: Scalable 3D Gaussian Diffusion Inference with Temporal and Spatial Trimming},
    author={Yin, Zeyuan and Liu, Xiaoming},
    booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS)},
    year={2025}
}