ViDiCalib — camera calibration from a handful of video frames

Release 2 (October 2026): learned ray fields meet prior-anchored bundle adjustment

Release 2 accompanies the paper Calibrating Cameras from a Handful of Video Frames: Learned Ray Fields Meet Prior-Anchored Bundle Adjustment. A multi-view ray-field network predicts a camera for a short clip; the camera anchors incremental mapping, enters bundle adjustment as a prior (with a low-rank prior shape) and verifies the result through a consistency gate. Code and checkpoints: chenming-wu/vidicalib-code (private).

path	contents
v2/RESULTS.md	the paper's numbers as tables (multi-view N = 2–10, six-camera test, single view)
v2/BENCHMARKS.md	benchmark definitions, protocol, metrics, development / test hygiene
v2/manifests/train/	training manifests: PanShot clips, panorama renders, single-view renders (frame paths, cameras, conditions)
v2/manifests/dev/, v2/manifests/test/, v2/manifests/newtest/	development, regression-set and new-test definitions; ADT as clip definitions only
v2/results/rows_node{0,1}.tar.gz	every per-clip result row of the project: all methods, benchmarks, N and seeds (runs/nview, runs/nview_fused, runs/single, COLMAP / DroidCalib scores, BA outputs, the new test)
v2/results/tables/	the paper's tables (LaTeX), the full comparison tables (P3_TABLES.md) and the decision outputs

Headline (N = 10, AE mean in degrees; two seeds):

	ADT-RGB	ADT gray	TUM-VI
AnyCalib (g), shared fit	1.44	1.21	1.79
DroidCalib (on its successful clips)	2.63 (73/100)	1.65 (75/100)	6.85 (2/69)
ours, full system	0.62	0.49	0.50

On six cameras that played no role in development (pre-registered, opened once) the full system reduces the network's error by 20% (95% CI 16–24%) and is 15% more accurate than AnyCalib with identical post-processing. Our network is not state of the art on single images. CalibAnyView's published numbers come from different ADT clips and are reference only. See v2/RESULTS.md. After an internal audit (2026-10-07) the evaluation and decision scripts were made strict about completeness and re-run; every reported number is unchanged (v2/results/verification/).

No third-party media are redistributed. ADT clips are given as definitions (sequence, camera, start frame, stride) because the Aria Digital Twin license does not permit redistribution; TUM, UZH-FPV and PanShot clips are given as manifests with their published calibrations and remain governed by those datasets' licenses (see ATTRIBUTION.md). Our panorama renders are reproducible from the manifests and the public panoramas.

Release 1 (September 2026): wide-angle camera calibration from video

Synthetic video clips rendered from equirectangular panoramas through known wide-angle camera models, with exact per-pixel ground truth: the ray field, gravity, latitude/up fields and validity masks used to produce the pixels are returned by the renderer, not estimated from them.

This release contains the labels, manifests, out-of-distribution test set and measured results. The rendered RGB for the 49,434-clip training split is 359 GB and is not included here; it is reproducible from the manifests and the public panoramas (see Reproducing the renders).

What is here

path	contents
manifests/paper.manifest.jsonl	49,434 clips, 4,004,154 frames, 749 panoramas
manifests/ood.manifest.jsonl	600 clips from measured lens profiles
manifests/paper.qc_report.json	QC gate result for the training split
labels/paper_dense_gt.tar	exact dense ground truth, one .npz per clip
labels/paper_poses.tar	camera trajectories
labels/ood_*.tar	the same for the OOD split, plus its shared ray grids
ood/ood_videos.tar	the OOD clips' RGB (3.7 GB) — usable as-is
lens_profiles/	6,474 Lensfun/Optiland profiles, 1,436 families, with the train/test family split
results/	the controlled three-arm experiment: per-clip metrics and summaries
docs/	conventions, data provenance, results, model card, repair report

Conventions

Camera frame is x right, y down, z forward. Pixel centres are at integers — AnyCalib and GeoCalib put them at integer + 0.5, so cx_ours = cx_theirs − 0.5. Gravity is g(roll, pitch) = (sin r cos p, cos r cos p, −sin p); GeoCalib's is the negation of this. Getting either wrong produces plausible-looking numbers that are wrong by a constant, so docs/CONVENTIONS.md states them explicitly.

Splits

Partitioned by panorama and by source group, never by clip, so no panorama contributes to two splits: train 44,220 / val 2,574 / test 2,640.

The OOD split is separate and stronger: its lenses come from 127 lens families the training split never contains, and its ground truth is each profile's own ray grid rather than a parametric fit — 88% of held-out profiles have no converged parametric fit, and those that do are off by up to 1°.

Imaging conditions

clean 0.40, sensor noise 0.12, night 0.12, low texture 0.12, motion blur 0.08, dynamic occlusion 0.16. Motion blur is a genuine exposure integral, so it smears along the actual motion. Occluders are accepted only under a recoverability invariant — every pixel visible in at least one frame — because a pixel hidden in every frame is recoverable by no method and only adds label noise. Measured worst case over the whole split: 0.001.

Results included

A controlled three-arm comparison of generative geometry pre-training, with every variable fixed except the pre-trained branch:

arm	test ray AE (2,640 clips)	OOD ray AE (600 clips)
no geometry pre-training	10.312	9.724
Stage-1a	10.431	10.879
Stage-1a + 1b	10.128	10.149

Paired bootstrap against the no-pre-training arm: Stage-1a is significantly worse (test +0.223°, CI [+0.133, +0.328]; OOD +0.965°, CI [+0.744, +1.170]); Stage-1a+1b is not distinguishable (test −0.087°, CI [−0.211, +0.039]; OOD +0.199°, CI [−0.212, +0.520]).

The pre-training does learn geometry — free-running generation improves ray angular error from 101° to 19° and temporal consistency from 60.9° to 0.68° — but that did not transfer to downstream calibration accuracy. For reference, AnyCalib scores 4.804° on the same test split.

Reproducing the renders

The 359 GB of training RGB is deterministic given the manifests and the public OpenPano panoramas. Each manifest record carries the camera model and parameters, the trajectory, the encoder settings, the imaging condition and a SHA-256 of the produced video, so a re-render can be checked byte for byte.

Limitations

Translation is ignored (a panorama is at infinity), so parallax is absent. Occlusion is procedural, not real objects. The Laval Indoor HDR panoramas the OpenPano split lists also name require a manual application and are absent — 2,162 of 2,908 listed panoramas are missing, so the set is outdoor- and architecture-heavy relative to what those lists describe.

See ATTRIBUTION.md for licensing: the panoramas and the lens database carry their own terms and those flow through to these derived clips.