docs/DATA.md

# Data card

Covers what this repository actually produced, and what the design doc calls for
but could not be obtained here. Design-doc sections 6, 8, 9, 13.

## Line A — exact synthetic panorama video

**Source.** OpenPano, the panorama collection assembled for GeoCalib and reused
by AnyCalib. Public HTTPS download from ETH:
`https://cvg-data.inf.ethz.ch/GeoCalib_ECCV2024/openpano.zip` (7.5 GB).

**What arrived.** 749 panoramas: 670 train / 39 val / 40 test, at 10000x5000 or
16k, in the authors' own panorama-level split. Provenance groups: 618
poly-haven-style `*_16k`, 130 `hdrmaps_com_free`, 1 other.

**What is missing.** The split lists name 2,908 panoramas. The remaining 2,162
are Laval Indoor HDR, which requires a manual application to the authors
(`tools/download/fetch.py info laval`). The indexer reports the shortfall per
split instead of silently skipping, and `select_panos` prints it on every run.
Consequence: the synthetic tier here is built from the outdoor/HDRI part of
OpenPano only, so indoor scenes are under-represented relative to the paper
configuration.

**Split policy.** The authors' panorama-level split is honoured rather than
re-drawn — re-splitting would silently invalidate the AnyCalib/GeoCalib
comparison, since those models were trained on this partition.

**Generated tiers.**

| tier | panoramas | clips | frames total | resolution | size | wall time |
|---|---|---|---|---|---|---|
| smoke | 20 | 200 | 16,200 | 832x480 | 732 MB | 285 s (12 workers) |
| MVP | 749 | 11,984 | 970,704 | 832x480 | 51 GB | 5.2 h (12 workers) |

MVP splits: 10,720 train / 624 val / 640 test clips from 749 distinct panorama
sources. Camera-model counts 4,065 pinhole / 4,008 Brown-Conrady / 3,911 EUCM
(0.339 / 0.334 / 0.326 against the 0.34 / 0.33 / 0.33 target). Trajectory
families 7,185 rotation-only / 3,514 smooth 6-DoF / 1,285 difficult. Zero split
leakage, zero manifest integrity errors, 1,169 clips (9.8%) flagged
`resolution_limited`.

Camera-model mixture reproduces AnyCalib's `openpano_v2_gen` config exactly:
pinhole 0.34 / `radial:1` 0.33 / EUCM 0.33, with `vfov ~ U(20, 105)` deg for the
first two and `U(50, 180)` deg for EUCM, `k1 = k1_hat * f / h` with
`k1_hat ~ truncnorm(+-0.3, sigma 0.07)`, `alpha ~ U(0.5, 0.8)`,
`beta ~ U(0.5, 2.0)`. On top of that ViDiCalib samples what the doc asks for and
AnyCalib's generator does not: 30% of clips have an off-centre principal point
(within +-10% of W/H) and 20% have non-square pixels (+-5%).

Trajectory mixture (doc section 7.1): rotation-only 0.34, smooth 6-DoF 0.29,
real-motion replay 0.27, difficult 0.09. **The real-motion share is a
fallback**: no CameraBench / TUM VI pose files were available here, so those
clips are sampled rotation-only trajectories tagged `fallback=True` with
`intended_family="real_replay"`. `pool_summary` reports both the actual and the
intended mixture so this cannot be misread.

**Per-clip storage.** One MP4 (H.264, CRF 14, preset medium), one Stage-1a
overlay MP4, one Stage-1b overlay MP4, and a compact NPZ. The NPZ stores only
what cannot be recomputed — camera parameters, validity mask, per-frame gravity
and rotations, non-trivial coverage — so it is ~8.5 KiB per clip instead of the
93 MiB a stored `81 x 480 x 832 x 3` float16 ray field would take. Latitude, up
field and the ray field are regenerated by `expand_dense_gt`.

**Resolution honesty.** A 480-row clip at 100 deg vertical FoV needs ~0.21
deg/px, i.e. a panorama of at least ~1,729 rows before supersampling. Panoramas
are loaded at 2,560 rows. Clips whose target camera is *narrower* than the
panorama can resolve are flagged `resolution_limited` in the manifest, with the
achieved ratio recorded in `generator.pano_resolution_ratio`. In the smoke tier
21 of 200 clips carry that flag.

## Line B — physical lens distortion

**Lensfun** (public git, `third_party/lensfun`). Parsed: **6,432 measured
distortion calibrations across 1,429 lens families** — `ptlens` 5,554, `poly3`
873, `poly5` 5. For scale, AlignDiff's LensView bank is 3,187 profiles and is
not downloadable.

Built bank (`/data/derived/lenses/profiles.jsonl`): 6,474 profiles including
Optiland, across 1,436 families.

| distortion family | profiles |
|---|---|
| pincushion | 2,732 |
| barrel | 2,203 |
| negligible | 1,326 |
| moustache | 177 |
| fisheye | 36 |

Near-duplicate clustering found **2,169 distinct radial signatures among the
6,474 profiles**, i.e. 4,255 are near-copies of another entry at the analysis
resolution. That is worth knowing before quoting the raw count as diversity.

Family split: 999 train families and 296 test families, with
`test_families_unseen_fraction = 1.0` -- every test family is absent from
training, against the doc's 20% requirement.

Brown-Conrady `k1` fitted from 120 real profiles spans `[-0.376, 0.221]`, which
brackets the range AnyCalib's synthetic `k1_hat ~ truncnorm(+-0.3)` produces;
`p1` and `p2` come out within `+-2e-4` of zero, because the open profile formats
are purely radial.

**Coverage gap: decentred / asymmetric distortion.** None of Lensfun's three
present models (`ptlens`, `poly3`, `poly5`) can express it -- they are radial in
one variable -- and the database contains no `acm` entries, the one format that
carries tangential terms. The v1 synthetic sampling does not cover it either,
because it follows AnyCalib's `radial:1` with `p1 = p2 = 0`. In this repository
asymmetric distortion is therefore represented only by the non-parametric
residual field, which handles it by construction but gives the model no
*training* signal for it. Closing this needs `acm`-format profiles, decentred
Optiland designs, or measured LUTs from the real rig.


Every profile is converted to the canonical `pixel -> unit ray` grid plus a
64x64 `ideal-pixel -> distorted-pixel` warp grid, with the Hugin-to-normalised
unit conversion taken from lensfun's own source rather than guessed. Per-profile
analysis records maximum displacement, monotonicity, Jacobian sign consistency,
valid fraction, effective image circle and fittable FoV. Splits are by lens
**family**, so the 24 mm and 70 mm nodes of one zoom never straddle a split.

**Optiland** (`pip install optiland`, MIT). Perturbed variants of its sample
designs, traced with real rays to get image height versus field angle, then
inverted. 42 valid profiles from 78 attempts across 7 designs (`cooke_triplet`,
`double_gauss`, `reverse_telephoto`, `telephoto`, `tessar`, `heliar`,
`aspheric_singlet`); `petzval` and `field_corrector` never traced successfully
and are reported as rejected rather than quietly dropped. Coverage is barrel and
pincushion only — the sample designs are rectilinear photographic objectives, so
fisheye and moustache come from Lensfun and from the synthetic
KB / EUCM / Double-Sphere sampling.

**LensView** (AlignDiff's commercial source) is not used and not required; the
open path above is the main line, exactly as doc section 8.1 plan B specifies.

## Line C — real 7V capture

**Not collected.** This requires vehicle access, the camera rig, and the
calibration workflow of doc section 9 (Kalibr / BabelCalib / Basalt, AprilGrid
sequences, IMU static segments, temperature sweeps). What exists here is the
*plumbing*: `ClipRecord` carries `domain_id`, `vehicle_id`, `route_id`,
`gt_confidence` and `gt_ray_sigma_rad`; `check_split_leakage` enforces the
instance / vehicle / route / firmware split policy of doc section 9.3; and
`SourceCalibration.ray_sigma` propagates the uncertainty of an *estimated*
source calibration so re-distorted real footage is never treated as exact GT.

## Registration-gated sources

Not obtained, and not scriptable. `tools/download/fetch.py info <name>` prints
the exact procedure for each.

| source | reason |
|---|---|
| Laval Indoor HDR | manual application to the authors |
| KITTI-360 | registration + non-commercial licence |
| WoodScape | proprietary licence |
| Aria Digital Twin | registration, then a per-user CDN JSON |
| CameraBench train split | may require a form |

## Public-but-not-fetched sources

`360-1M`, `UCPE`/`PanShot`, `CO3Dv2` and `MegaDepth`/`TartanAir` have working
fetchers in `tools/download/fetch.py` but were not downloaded here. For 360-1M
the fetcher deliberately stops after writing a filtered pilot selection: the
full set is ~200 TB, the filtered 24k subset ~5 TB, the YouTube URLs decay, and
the licence has to be checked per source.

## Licences

* OpenPano — per-panorama; the poly-haven and hdrmaps portions are the
  permissively licensed part of the collection. Check the upstream terms before
  redistribution.
* Lensfun — CC-BY-SA / GPL depending on component; the profile *data* is an open
  database.
* Optiland — MIT.
* Wan2.1-VACE-1.3B — Apache-2.0.
* AnyCalib, GeoCalib checkpoints — see their repositories.
* CO3Dv2 — CC BY-NC 4.0 (non-commercial).

Raw downloads live outside the repository (`/data/raw`, `/data/models`) and are
referenced by configuration. No data, weights or cookies are committed.