README.md

metadata

license: cc-by-nc-4.0
pretty_name: IDSet-3D
tags:
  - 3d-reconstruction
  - depth-estimation
  - benchmark
  - evaluation

IDSet-3D data

Benchmark labels, ID-Head features, model predictions, and the outputs of the registered formal run of IDSet-3D, an identified-set evaluation toolkit for 3D reconstruction. The evaluator, the scripts, and the documentation are in the IDSet-3D GitHub repository, whose DOWNLOADS.md also lists the upstream data, weights, and code with the versions used.

The folders mirror the layout of the repository root, which the scripts use as their data root. Download into a clone of the GitHub repository, all at once or by folder:

pip install -U huggingface_hub
python - <<'EOF'
from huggingface_hub import snapshot_download
skip = ["README.md", ".gitattributes"]   # keep the GitHub repository's README.md and .gitattributes
snapshot_download("wolfkingttf/IDSet3D", repo_type="dataset", local_dir=".", ignore_patterns=skip)
EOF
sha256sum -c checksums/IDSet3D.sha256 --ignore-missing

Skipping README.md and .gitattributes keeps the files of the GitHub repository from being replaced. To fetch only some folders, add allow_patterns=["predictions/*", "checksums/*"] to the call.

Contents

Ray files cover the 90×160 evaluation grid of one query view (14,400 rays); those in pool/dev_150m/ use a 180×320 grid. Block 1 holds 12 windows: the 10 test-1 windows and the 2 development windows.

Folder	Files	Size	Content
eval/eval-v1.0/oracle/matrixcity/	1,046	208 MB	ray labels of the 12 block-1 windows for every condition: upper bound b, uncertain length u, range L, tolerance tau, endpoint state (V/H/N/X), edge flag, and true first-hit distance d_gt
pool/*.json	9	2.4 MB	windows and query views of the training, validation, and test-2 splits, with their provenance
pool/query_v2/	2,054	395 MB	ray labels of the 1,562 training, 368 validation, and 123 test-2 query views, and the query selection
pool/c76_labels_h_native/	1,930	365 MB	ID-Head training targets of the training and validation queries: answered rays, endpoint states, front and back violations
pool/features/VGGT-1B/	1,930	1.05 GB	the 18 per-ray ID-Head features computed from VGGT-1B, with the readout range and confidence, for the training and validation queries
predictions/	14,930	687 MB	predictions of the eight models on the 12 block-1 windows under all registered conditions (predictions/<model>/matrixcity/<window>/<condition>/<query>__run0.npz: range in meters and confidence per ray)
eval/c83_formal/, eval/c83_method_work/, eval/c86_*/	3,003	427 MB	outputs of the registered formal run: test-2 labels and predictions of all models; ID-Head features, labels, and scores on both test sets; checksums and logs; the independent recomputation
eval/v1.0/full/	54	18 MB	test-1 benchmark results (read by release/reproduce_main_tables.sh)
pool/raw/, eval/c83_formal/test2/raw/	264	6.0 GB	raw model outputs on the input views (point maps, confidence, cameras), for recomputing features or predictions without a GPU
pool/dev_150m/, dev_native/	264	106 MB	labels, features, and VGGT-1B raw outputs of the two development windows, on a 180×320 grid (pool/dev_150m/) and on the 90×160 grid (dev_native/)

Not included: the MatrixCity images and depth maps (download them from the official release; the formal-run folders frames/ and reselect_depth/ were copies of them), the model weights, and the predictions on the Asset Inspection scenes, which enter no reported result. A void 3DGS run with random initialization is excluded, as in the GitHub repository.

Checksums

checksums/IDSet3D.sha256 covers every file. The GitHub repository records the same SHA256 for the predictions (release/predictions_manifest.json, paths relative to predictions/) and for the formal-run outputs (release/results_v2_manifest.json, absolute paths under /root/gpufree-data/IDSet3D/). Three result files of eval/v1.0/full/ differ from their frozen copies in one note each, and one of them in its file name: gtpose_secondary.json, robust_control_v2.json, and twin_m5_v2.json (see release/RELEASE_EDITS.md in the GitHub repository). No value was changed.

License

CC BY-NC 4.0. The files are derived from MatrixCity, which is released under CC BY-NC 4.0. The predictions and raw outputs were produced by the models listed in DOWNLOADS.md of the GitHub repository, and the terms of those models may also apply.