MANIFEST.json

{
  "tartanair2": {
    "path": "data/ext/tartanair2",
    "size_gb": 73,
    "license": "CC BY 4.0",
    "redistribute": "yes",
    "attribution": "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.",
    "note": "12 per-environment zips, read in place by src/pointcalib/data/tartanair.py; 105,922 lcam_front frames."
  },
  "hypersim": {
    "path": "data/ext/hypersim",
    "size_gb": 7,
    "license": "CC BY-SA 3.0",
    "redistribute": "yes",
    "attribution": "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.",
    "note": "14,505 depth_meters.hdf5 + tone-mapped jpgs, 154 scenes."
  },
  "openscene": {
    "path": "data/ext/openscene",
    "size_gb": 12,
    "license": "CC BY-NC-SA 4.0 + Motional ToU",
    "redistribute": "yes",
    "attribution": "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.",
    "note": "cam_0/lidar_0/metadata tgz + openscene-v1.1 + our 1,943-frame derived cache (own geometry pipeline; inherits NC + share-alike)."
  },
  "lightwheelocc": {
    "path": "wds/lightwheelocc-*.tar",
    "size_gb": 7.4,
    "license": "CC BY-NC-ND 4.0 upstream + separate authors' permission for this repo",
    "redistribute": "by permission only -- NOT transferable to downstream users",
    "attribution": "LightwheelOcc (Lightwheel AI), OpenDriveLab/LightwheelOcc. Upstream is CC BY-NC-ND 4.0. The ND clause bars distributing a derivative, and re-encoded shards ARE a derivative, so the public licence alone does not permit this split. It is published under separate permission granted by the dataset's authors to this repository's owner. That grant is specific to this repository and does NOT travel with the files: anyone redistributing wds/lightwheelocc-*.tar onward needs their own permission from Lightwheel AI. Recorded this way deliberately rather than as 'redistribute: yes', because a private grant to us is not a public licence for third parties.",
    "note": "40,800 frames in 21 shards: 34,000 samples x 6 cameras = 204,000 raw frames, taken at stride 5 (0.5 s spacing; consecutive samples are ~0.1 s apart and six synchronised cameras already give six near-independent views). Verified conventions across 144 frames and all six camera mounts: depth is a 3-channel uint8 PNG carrying a 16-bit CENTIMETRE value as ch0 + ch1*256 with ch2 identically zero (unique values per channel 256/233/1); PLANAR z-depth (ground-plane residual 14.3 units vs 39.5 euclidean); 0 means invalid and is how sky is stored (18.1%), while 65535 -- the 16-bit ceiling at 655.35 m -- essentially never occurs, so there is no saturated far sentinel. The centimetre scale was recovered from the dataset's OWN metadata rather than guessed: sensor2ego_translation[2] gives each camera's mounting height, and a robust ground-plane fit agrees within 1.6% for five of the six (CAM_BACK_LEFT is the exception because its lower image is kerb, not road). A 24-bit CARLA-style (R + G*256 + B*65536) unpack decodes without error and is WRONG -- it puts median depth at 496 m. Intrinsics are per camera AND per frame from lightwheel_occ_infos_*.pkl and are carried through verbatim; note CAM_BACK is genuinely wider than the rest (fx 809 against ~1255), which is real intrinsic diversity rather than an error. 94.2% of usable pixels fall within 80 m."
  },
  "matrixcity": {
    "path": "wds/matrixcity_*.tar",
    "size_gb": 12.7,
    "license": "CC BY-NC 4.0",
    "redistribute": "yes",
    "attribution": "MatrixCity (Li et al., ICCV 2023), BoDai/MatrixCity, CC BY-NC 4.0 -- attribution, non-commercial, redistribution granted. The aggregate corpus is already non-commercial (OpenScene/nuPlan, KITTI), so this adds no new restriction.",
    "note": "62,452 frames in 38 shards across 12 independently streamed chunks (both cities x 8 street-camera regimes). Verified conventions, all measured before the upstream docs were read and all four then matching them: depth_m = value/100 (CENTIMETRES -- independently corroborated by the 0.01 norm of transforms.json's rot_mat rotation rows, i.e. the conversion is baked into the pose matrix); PLANAR z-depth (ground-plane residual 11.6 units vs 25.7 euclidean); 65504 is the SKY SENTINEL rather than a distance, being float16's maximum, so 655.04 m is the farthest representable value; optical axis -z (NeRF X-right/Y-up/Z-backward, 87% two-view reprojection inliers vs 27% for +z); camera_angle_x = pi/2 exactly so fx = fy = W/2. Only `street` is included (aerial is a different task; the normal/diffuse/roughness/specular/metallic passes are irrelevant to depth), and only test + train, not train_dense, and for big_city only test: 118.9 GB of the 824 GB street split. That is a diversity-per-GB choice -- our scaling curve responded to new DOMAINS while frames WITHIN a domain saturated at ~1k, train_dense is the same trajectories sampled more densely in time, and big_city/street/train is 410 GB of the same city as big_city/street/test. Frames under 20% valid coverage are dropped (3.4% overall, 18% in the sky-heavy small_outside_test chunk). This split is also why encode_depth now MASKS beyond 255.99 m instead of clipping: 1.3% of MatrixCity's valid pixels lie past it and clipping would have built a fabricated wall of geometry at exactly that range."
  },
  "urbansyn": {
    "path": "wds/urbansyn-*.tar",
    "size_gb": 1.4,
    "license": "CC BY-SA 4.0",
    "redistribute": "yes",
    "attribution": "UrbanSyn (urbansyn.org), which states the dataset is released under CC BY-SA 4.0 -- attribution + share-alike, commercial use permitted, redistribution explicitly granted.",
    "note": "7,539 frames. Verified conventions: depth_m = EXR value * 1e5 (the site documents '1e5 meters' and carries a changelog correcting an earlier '10e-5' claim); PLANAR z-depth (one-plane RANSAC favours z in 10/12 frames, mean explainable fraction 48.2% vs 43.8%); fx = fy = 1730.21, c = (1024, 512) from the site's camera_metadata.json, cross-checked against its own fov_deg 61.237. Sky is masked by SEMANTIC LABEL (Cityscapes train id 10), not by a depth threshold -- the per-frame far value was measured at 354-580 m across six frames, so no fixed cutoff separates sky from real distant geometry. Sampled shard: 38 of 50 frames have <99% valid pixels, confirming the mask engages."
  },
  "pointodyssey": {
    "path": "wds/pointodyssey-*.tar",
    "size_gb": 1.9,
    "license": "MIT",
    "redistribute": "yes",
    "attribution": "PointOdyssey (Zheng et al., ICCV 2023), MIT licensed on aharley/pointodyssey. The only component of this corpus with unambiguous redistribution terms.",
    "note": "13,519 frames from the val + test splits, every 5th frame (sequences run 849-2750 frames at 30 fps, so consecutive frames are near-duplicates). Verified conventions against the dataset's OWN 3D point trajectories -- the strongest anchor available anywhere in this corpus: projecting anno.npz trajs_3d through extrinsics and dividing the depth reading by camera-space z gives 65.16 over 845,095 points from 302 frames, with 0.35% cross-frame spread, versus 2.81% for the euclidean reading (8.0x worse). depth_m = value / 65535 * 1000 is 0.569% off the fitted constant and is adopted as the round encoding. Two traps: info.npz stores SHAPES rather than data (its 'intrinsics' entry is literally [849, 3, 3]) so anno.npz is authoritative, and scene_info.json's sensor_width differs per scene (50.0 vs 36.0 in two sampled scenes)."
  },
  "mvssynth": {
    "path": "wds/mvssynth-*.tar",
    "size_gb": 2.1,
    "license": "none stated; \"research and educational use only\"",
    "redistribute": "unclear",
    "attribution": "MVS-Synth, from DeepMVS (Huang et al., CVPR 2018), rendered from Grand Theft Auto V. phuang17.github.io/DeepMVS/mvs-synth.html states only: \"The data is for research and educational use only.\" There is NO license text and NO explicit grant to redistribute, and the HF mirror phuang17/MVS-Synth tags itself license:other. Two unresolved layers: the authors grant no redistribution right, and the underlying frames derive from Take-Two/Rockstar game assets whose EULA we have no grant under either. Upstream is itself publicly downloadable without gating, and what we publish is a downsampled 640x360 derivative, but neither fact is a redistribution licence. FLAGGED for the repo owner to decide; removing wds/mvssynth-*.tar deletes it without touching anything else.",
    "note": "12,000 frames, 120 sequences. Verified conventions: depth is DECIMETRES (0.1 m/unit, from three independent absolute-scale anchors), planar z-depth, sky = inf, per-sequence f_x 530.9-578.7. All normalised to metres + per-frame K in the shards."
  },
  "eval_protocol": {
    "path": "runs/shared_kitti_val291",
    "size_gb": 1.6,
    "license": "CC BY-NC-SA 3.0 (KITTI)",
    "redistribute": "yes",
    "attribution": "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.",
    "note": "rgb + velodyne_raw sparse + cleaned GT + K, and pred_*.npy for OMNI-DC, Marigold-DC, PromptDA, PriorDA, MoGe-2, DAv2 and ours."
  },
  "eval_shared60": {
    "path": "runs/shared_kitti",
    "size_gb": 1.1,
    "license": "CC BY-NC-SA 3.0 (KITTI)",
    "redistribute": "yes",
    "attribution": "Same as eval_protocol; the earlier 60-frame shared set.",
    "note": "kept for continuity with the tables computed on it."
  },
  "checkpoints": {
    "path": "runs/scale120k",
    "size_gb": 0.06,
    "license": "ours",
    "redistribute": "yes",
    "attribution": "Trained by us; no third-party data redistributed.",
    "note": "121k-corpus run, step-tagged snapshots 6k..60k. The 42k step is the train-split-selected checkpoint behind the reported numbers."
  },
  "checkpoints_base": {
    "path": "runs/nyu_v1",
    "size_gb": 0.11,
    "license": "ours",
    "redistribute": "yes",
    "attribution": "Trained by us.",
    "note": "the original NYU-only baseline checkpoint."
  },
  "recipe": {
    "path": "__recipe__",
    "size_gb": 0.01,
    "license": "ours",
    "redistribute": "yes",
    "attribution": "Ours.",
    "note": "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."
  }
}