EFM3D: A Benchmark for Measuring Progress Towards 3D Egocentric Foundation Models
Paper • 2406.10224 • Published
These weights are Meta's, redistributed here only as a convenience mirror for the Fire3D perception baselines. Your use of them is governed by the original Project Aria licenses, not by this repository.
Fire3DBaselineWeight
A convenience mirror of the two gated baseline checkpoints used by Fire3DPerceptBaseline — EVL (EFM3D) and SceneScript.
No weights here were trained by this repository. Both models are Meta's,
originally distributed from projectaria.com behind an email + license-acceptance
form that issues per-submission, expiring download links. That makes them
impossible to fetch from a setup script, which is the only reason this mirror
exists. The files are byte-identical to what that form produced; they are simply
already extracted from evl_model_ckpt.zip.
Contents
| file | size | sha256 | source |
|---|---|---|---|
efm3d/model_release.pth |
426 MB | 9638b6e0127ba1962e267f01ba5c310e42d8f22a8a5a2fee65de0bb64752ec1e |
EVL / EFM3D, joint 3D detection + surface regression. Needs ~20 GB VRAM. |
efm3d/model_lite.pth |
424 MB | 61326af5568abe9cf892eaa3fbe29d708357f9e7114f369d4f2878379ec6d44f |
EVL / EFM3D, desktop-GPU variant. |
scenescript/ase_model.ckpt |
114 MB | 486f43d6bd55d7b341aa7e18394d22f93ecbd47b8e3004eb59f30f2866fdf824 |
SceneScript, ASE model — predicts layout and classed boxes. |
The seq136_sample smoke-test sequence that ships inside evl_model_ckpt.zip is
not mirrored here.
Not mirrored either, because they already have public scriptable sources:
BoxerNet / DinoV3 / OWLv2 (facebook/boxer)
and the DinoV2 ViT-B/14 EVL backbone (dl.fbaipublicfiles.com).
Use
From a Fire3DPerceptBaseline checkout, everything lands in the right place with:
bash scripts/download_ckpts.sh
Or directly:
hf download TianhangCheng7/Fire3DBaselineWeight \
efm3d/model_release.pth --local-dir .
Provenance and licensing
Copyright and all rights in these checkpoints remain with Meta Platforms, Inc. They are covered by the license terms presented on the Project Aria download form, and mirroring them here neither grants nor implies any additional rights. If you intend to use them for anything beyond reproducing these baselines, get them from the pages above and read the terms yourself. Meta may request removal of this mirror at any time, in which case it will be taken down.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support