You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These are ScanNet v2 mesh files, redistributed here only to make a specific evaluation reproducible. Access requires that you have ALREADY signed the ScanNet Terms of Use with the ScanNet authors.
Log in or Sign Up to review the conditions and access this dataset content.
thinking-in-space — dense meshes for the official_val_v2 eval
The 68 ScanNet scenes used by thinking_in_space_official_val_v2 (n=256 QA), shipped as the
dense *_vh_clean.ply (~2 M vertices each, 14.2 GB total).
Why this exists. Our eval renders scenes from the mesh at run time, and the coarse
_vh_clean_2.ply(~80 k verts) costs ≈0.12 accuracy versus the dense mesh — with vertex-colour shading, vertex density is the texture resolution. Every number in our results tables is dense-mesh. A server holding only the coarse meshes will silently produce non-comparable results. The public data bundle (Icey444/tmp : 3d-train-full/scans) contains coarse meshes only, which is the gap this repo closes.
⚠️ ScanNet Terms of Use
This is ScanNet v2 data, not ours. It is redistributed only for evaluation reproducibility and is gated: you must have signed the ScanNet ToU with the ScanNet authors before requesting access, and you may not redistribute it further. If you already have ScanNet locally, use your own copy — you do not need this repo.
Contents
scans/<scene_id>/<scene_id>_vh_clean.ply # 68 dense meshes
eval_scene_ids.json # the 68 scene ids
frame_poses.json # start pose per scene (eval spawns here)
scene_bounds.json # per-scene world bounds (prompt hint + clipping)
Use
export HF_TOKEN=<your token>
python - <<'PY'
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Icey444/tis-eval-dense-meshes", repo_type="dataset",
local_dir="eval_meshes", max_workers=4)
PY
TIS_VARIANT=<v9|v9.1> MODEL=<merged-bf16-ckpt> \
EVALSET=<official_val_v2/eval.parquet> \
TIS_SCANS=$PWD/eval_meshes/scans TIS_DENSE_MESH=1 TIS_VAL_DUMP=dump \
CUDA_VISIBLE_DEVICES=0 bash projects/thinking_in_space/scripts/eval.sh
python projects/thinking_in_space/scripts/score_dumps.py dump
TIS_SCANS overrides the absolute mesh_path baked into the eval parquet, so you do not need the
original directory layout. Scoring needs a Qwen3-30B judge on :30000 (scripts/serve_judge.sh).
Code: williamium3000/verl-projects, branch
thinking-in-space-clean-pr → projects/thinking_in_space/.
Training data is elsewhere and needs none of this — the trajectory-SFT banks ship pre-rendered
frames (Icey444/tmp : zeroshot-trajs/*). These meshes are only for evaluation.
- Downloads last month
- 28