ChenmingWu/pointcalib-corpus

PointCalib corpus mirror

Training corpus and frozen evaluation protocol for PointCalib, a sparse-point-conditioned residual calibration field over a frozen geometry foundation model. Mirrored so results are reproducible bit-for-bit.

Components

component	size	license	redistribution
tartanair2	73 GB	CC BY 4.0	yes
hypersim	7 GB	CC BY-SA 3.0	yes
openscene	12 GB	CC BY-NC-SA 4.0 + Motional ToU	yes
eval_protocol	1.6 GB	CC BY-NC-SA 3.0 (KITTI)	yes
eval_shared60	1.1 GB	CC BY-NC-SA 3.0 (KITTI)	yes
checkpoints	0.06 GB	ours	yes
checkpoints_base	0.11 GB	ours	yes
recipe	0.01 GB	ours	yes

Attribution and terms

tartanair2

TartanAir V2, AirLab / CMU. tartanair.org states: "The TartanAir V2 dataset is licensed under a Creative Commons Attribution 4.0 International License" -- attribution only, no share-alike, commercial use permitted. NOTE: the HF mirror theairlabcmu/tartanair2 tags itself bsd-3-clause, which is the license of the castacks/tartanair_tools CODE, not the dataset; we carry the upstream CC BY 4.0 instead.

12 per-environment zips, read in place by src/pointcalib/data/tartanair.py; 105,922 lcam_front frames.

hypersim

Hypersim, Apple (Roberts et al., ICCV 2021). apple/ml-hypersim README: "The Hypersim Dataset is licensed under the Creative Commons Attribution-ShareAlike 3.0 Unported License" -- attribution + share-alike. (The repo LICENSE.txt is a separate Apple SOFTWARE license and does not govern the data.) The release excludes the purchased Evermotion source meshes; Apple does not address whether the upstream asset EULA independently permits redistribution of derived renders, so that residual question is unresolved rather than cleared.

14,505 depth_meters.hdf5 + tone-mapped jpgs, 154 scenes.

openscene

OpenScene-v1.1 (OpenDriveLab) over nuPlan (Motional). The nuScenes/nuPlan Terms of Use place the data under CC BY-NC-SA 4.0, which grants the right to "reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only"; the ToU adds no-endorsement, termination and indemnification terms but no redistribution prohibition. OpenScene itself self-describes as "a compact redistribution of the large-scale nuPlan dataset". Carry-over caveats: third-party-supplied portions may not be redistributed without the original provider's consent, and Motional may terminate access at any time.

cam_0/lidar_0/metadata tgz + openscene-v1.1 + our 1,943-frame derived cache (own geometry pipeline; inherits NC + share-alike).

eval_protocol

KITTI depth completion (Geiger et al.), CC BY-NC-SA 3.0: attribution, non-commercial, share-alike. This is the 291-frame day-disjoint val split used for every reported number, with baseline per-pixel predictions alongside.

rgb + velodyne_raw sparse + cleaned GT + K, and pred*.npy for OMNI-DC, Marigold-DC, PromptDA, PriorDA, MoGe-2, DAv2 and ours._

eval_shared60

Same as eval_protocol; the earlier 60-frame shared set.

kept for continuity with the tables computed on it.

checkpoints

Trained by us; no third-party data redistributed.

121k-corpus run, step-tagged snapshots 6k..60k. The 42k step is the train-split-selected checkpoint behind the reported numbers.

checkpoints_base

Trained by us.

the original NYU-only baseline checkpoint.

recipe

Ours.

download + preprocessing scripts, dataset manifests, configs, the 32-GPU plan, and the measured-findings summary, so the corpus can be rebuilt from upstream sources without relying on this mirror.

Excluded from this mirror

• sunrgbd (45 GB, NONE STATED): redistribute=no. EXCLUDED: single 45 GB zip, 82,692 entries, and it was never used in any training run, so excluding it costs the corpus nothing.

Rebuild these from their upstream sources with scripts/download_ext.py.

Measured results this corpus produced

Zero-shot KITTI, day-disjoint 291-frame val split, identical frames and masks for every method, paired Wilcoxon + 1e4 bootstrap:

metric	PointCalib	OMNI-DC
RMSE	1.1389	1.1220 (tie, p=0.194)
MAE	0.2800	0.2387
absrel	0.0177	0.0144
EdgeCR	68.42	59.72

PointCalib ties on RMSE and wins edge-structure preservation; it loses MAE and absrel. docs/PLAN_32GPU_SOTA.md in recipe/ documents the oracle upper bounds that localise why, and what a 32-GPU run would have to change.

Rebuilding without this mirror

uv run python scripts/download_ext.py          # tartanair2, hypersim, sunrgbd
uv run python scripts/export_shared_eval.py --data kitti_selection \
    --kitti-split val --limit 291 --out runs/shared_kitti_val291