Datasets:

Icey444
/

tis-train-dense-meshes

Modalities:

Size:

Tags:

Libraries:

License:

Dataset card Data Studio Files Files and versions

xet

Community

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 only to make a specific training/eval setup reproducible. Access requires that you have ALREADY signed the ScanNet Terms of Use with the ScanNet authors.

thinking-in-space — DENSE meshes for the training scenes

The 630 ScanNet training scenes used by Icey444/tmp : 3d-train-full, shipped as the dense *_vh_clean.ply (~184 MB/scene, 116 GB total).

Why this exists. The main data bundle (Icey444/tmp : 3d-train-full/scans) ships the coarse _vh_clean_2.ply only (~4.7 MB/scene). Dense is 32× larger — which is why it was left out — but with vertex-colour shading, vertex density is the texture resolution, and the coarse mesh costs ≈0.12 accuracy. Every number in our results tables is dense-mesh, so a box holding only the coarse meshes silently produces non-comparable results.

Companion repo for the 68 eval scenes: Icey444/tis-eval-dense-meshes. The two scene sets are disjoint (train vs official_val_v2).

⚠️ ScanNet Terms of Use

This is ScanNet v2 data, not ours. It is redistributed for reproducibility only and is gated: sign the ScanNet ToU with the ScanNet authors before requesting access, and do not redistribute further. If you already have ScanNet locally, use your own copy — you do not need this repo.

Layout

scans/<scene_id>/<scene_id>_vh_clean.ply     # 630 dense meshes

Matches the layout of the coarse bundle, so it can be merged over it, or pointed at directly.

Use

export HF_TOKEN=<your token>
python - <<'PY'
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Icey444/tis-train-dense-meshes", repo_type="dataset",
                  local_dir="dense", max_workers=4)     # keep workers low: HF API quota
PY

# then point the renderer at it
export TIS_SCANS=$PWD/dense/scans TIS_DENSE_MESH=1

TIS_DENSE_MESH=1 also makes verl/tools/render_view_tool.py auto-upgrade any _vh_clean_2.ply path to its dense sibling when present.

Training on pre-rendered trajectory banks does not need these meshes — the SFT banks ship their frames. Meshes are needed to render new viewpoints: RL rollouts, re-rendering trajectory frames, and evaluation.

Downloads last month: 46

Total file size:

116 GB