You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is derived from ScanNet and from annotations built on it (ScanRefer, Multi3DRefer, ScanQA, SQA3D, Grounded-3D-LLM proposals), and, for the LAM3C features only, from 3RScan and MultiScan scans (via SceneVerse). The scene files contain per-point coordinates of these scans. Access is granted only to researchers who have agreed to the ScanNet Terms of Use (http://kaldir.vc.in.tum.de/scannet/ScanNet_TOS.pdf) and, for the 3RScan and MultiScan scenes, to the 3RScan Terms of Use and the MultiScan license, for non-commercial research. Requests are reviewed manually.
Log in or Sign Up to review the conditions and access this dataset content.
just3d data
Inputs for just3d (private code repository):
per-point dense features of 1193 ScanNet scenes from six 3D encoders, object mask
proposals, and the training annotations of the reproduced runs (anypoint 1-61..1-84 /
2-92..2-115). LAM3C features (lam3c_b, lam3c_l) additionally cover all 1513
ScanNet scenes and the 3RScan / MultiScan evaluation scenes. The repository root mirrors data/ in the code repository, so a
download goes straight into it.
Layout
| Path | Files | Size | Contents |
|---|---|---|---|
annotations/ |
7 JSON + PROVENANCE.json |
0.2 GB | Training annotations, one file per mask source (*_clasp.json, region caption *_gt.json); PROVENANCE.json records each source file and its sha256. |
scene/mask_proposal/clasp/ |
563 | 1.0 GB | Grounded-3D-LLM (CLASP) top-100 proposals, packed masks over the full point set. |
scene/mask_proposal/gt/ |
1193 | 0.8 GB | ScanNet ground-truth instance masks. |
scene/dense_feature/<encoder>/ |
1193 each | 1.76 TB total | feat (bf16 [M, C]), coord_down ([M, 3]), inverse ([N]) per scene, for concerto_l (390 GB), utonia (314 GB), sonata (279 GB), concerto_b (279 GB), mask3d (190 GB), clasp (190 GB). |
scene/dense_feature/lam3c_b/, lam3c_l/ |
1802 each | 0.93 TB total | Same format, from LAM3C (aist-cvrt/lam3c-roomtours, PTv3 Base / Large; C = 1232 / 1728) for 1513 ScanNet scenes, 284 3RScan scenes (Beacon3D + 3RScan scene captioning, named by 3RScan UUID) and 5 MultiScan scenes (Beacon3D, scene_xxxxx_xx): lam3c_b (387 GB), lam3c_l (542 GB). Their inverse differs from the other encoders' by a few voxel-boundary points; read each encoder's own. |
The formats are specified in data/README.md of the code repository. Converted on
2026-10-01 from the anypoint feature stores; every object was verified against the
source store during conversion.
Download
hf auth login # after your access request is approved
hf download TorridFish/just3d-data --repo-type dataset --local-dir data
# one encoder only:
hf download TorridFish/just3d-data --repo-type dataset --local-dir data \
--include "annotations/*" --include "scene/mask_proposal/*" \
--include "scene/dense_feature/concerto_l/*"
License and terms
Research use only. The scene data derives from ScanNet and
is subject to the ScanNet Terms of Use; the 3RScan and MultiScan scenes of
lam3c_b / lam3c_l derive from 3RScan and
MultiScan (point clouds from
SceneVerse) and are subject to their terms; the annotations additionally follow the
licenses of ScanRefer, Multi3DRefer, ScanQA, SQA3D and Grounded-3D-LLM. Do not
redistribute.
- Downloads last month
- 80