SOTA Leaderboard¶
Calibrex may call a calibration method state of the art only for a claim that a
frozen audit marks supported. This page lists every audited claim, one
standing per sensor pair, and the target pairs that have no audited claim yet.
Current standings¶
As of 2026-10-02, two pairs have supported SOTA claims:
imu-lidar(Livox MID360, against LI-Init):- rotation and clock offset;
- the full extrinsic, second round. The first full-extrinsic round was refuted and stays listed.
lidar-vehicle(KITTI): the motion-only LiDAR-to-vehicle rotation, against KITTI's calib_imu_to_velo, on eight unseen drives.
See MID360 IMU-LiDAR Calibration and
KITTI LiDAR-Vehicle Rotation
for the scope of each claim and what it does not claim. Every other target
pair has a native method but no audited claim yet, except lidar-lidar, whose
pre-registered audit was refuted
(NTU VIRAL LiDAR-LiDAR),
and camera-imu, whose pre-registered audit was also refuted (6/7 gates; the
constrained gate failed on the shortest recording,
Hilti camera-IMU).
| Pair | Standing | Supported | Refuted | Incomplete | Target |
|---|---|---|---|---|---|
camera |
no_claim | 0 | 0 | 0 | yes |
camera-imu |
refuted | 0 | 1 | 0 | yes |
camera-lidar |
no_claim | 0 | 0 | 0 | yes |
gnss-imu |
no_claim | 0 | 0 | 0 | yes |
imu-lidar |
supported | 2 | 1 | 0 | yes |
imu-vehicle |
no_claim | 0 | 0 | 0 | yes |
ins-lidar |
no_claim | 0 | 0 | 0 | yes |
lidar-lidar |
refuted | 0 | 1 | 0 | yes |
lidar-vehicle |
supported | 1 | 0 | 0 | yes |
lidar-wheel_odometry |
no_claim | 0 | 0 | 0 | yes |
| Pair | Category | Quantities | Verdict | Gates (achieved/total) | Protocol |
|---|---|---|---|---|---|
camera-imu |
targetless_camera_imu | rotation, time_offset | refuted | 6/7 | hilti-camera-imu-v1 |
imu-lidar |
targetless_imu_lidar | rotation, translation, time_offset | refuted | 0/8 | mid360-imu-lidar-extrinsic-vs-li-init-v1 |
imu-lidar |
targetless_imu_lidar | rotation, translation, time_offset | supported | 4/4 | mid360-imu-lidar-extrinsic-vs-li-init-v2 |
imu-lidar |
targetless_imu_lidar | rotation, time_offset | supported | 4/4 | mid360-imu-lidar-rotation-vs-li-init-v1 |
lidar-lidar |
targetless_lidar_lidar | rotation, translation | refuted | 2/4 | ntu-lidar-lidar-v1 |
lidar-vehicle |
targetless_lidar_vehicle | rotation | supported | 3/3 | kitti-lidar-vehicle-v1 |
The machine-readable version is
sota_leaderboard.json
(slac.sota_leaderboard/v0.1).
What a standing means¶
| Standing | Meaning |
|---|---|
supported |
At least one audited claim for this pair meets every frozen required gate and its dataset-family and independent-rig coverage. |
refuted |
No claim is supported, and at least one required gate of an audited claim is contradicted by its evidence. |
incomplete |
Claims exist, but evidence or coverage is missing, so nothing may be claimed. |
no_claim |
A target pair without any audited claim. |
The counts for each pair are kept separate, so a refuted claim is never hidden behind a supported one. The leaderboard copies verdicts from the audits and never computes its own.
Auditing a claim¶
A claim is scoped by its modalities, the quantities it estimates, and a method category. For example, with illustrative thresholds:
schema_version: slac.sota_audit_protocol/v0.1
protocol_id: kitti-ins-lidar-hand-eye-v1
scope:
modalities: [ins, lidar]
quantities: [rotation, translation]
category: trajectory_hand_eye
claim_text: >-
Native INS-LiDAR hand-eye calibration matches the KITTI vendor calibration
within 0.2 deg and 5 cm (p95) across two sequences.
minimum_dataset_families: 2
minimum_independent_rigs: 1
requirements:
- requirement_id: kitti-rotation-p95
phase: phase2
description: p95 rotation error to the vendor calibration
dataset_family: kitti_raw
independent_rig: true
rig_id: kitti-2011-09-26
evidence_path: kitti-ins-lidar-benchmark.yaml
evidence_sha256: <SHA-256 of the benchmark artifact>
method_id: calibrex_native_hand_eye
metric: rotation_error_deg
statistic: metric_p95
comparison: less_equal
threshold: 0.2
provenance:
generator: manual
generator_version: "1"
Each piece of evidence is a digest-pinned slac.benchmark/v0.1 artifact whose
provenance declares data_verified: true. Freeze the protocol before looking
at the results, then run:
calibrex sota audit protocol.yaml --output audit.yaml
The command exits with 0 only for a supported verdict.
calibrex camera-lidar audit-sota keeps accepting Camera-LiDAR protocols, and
both commands evaluate claims with the same engine.
Regenerating this page¶
calibrex sota leaderboard AUDIT... \
--target-pair camera-lidar --target-pair lidar-lidar \
--target-pair imu-lidar --target-pair camera-imu --target-pair camera \
--target-pair gnss-imu --target-pair ins-lidar \
--target-pair lidar-vehicle --target-pair imu-vehicle \
--target-pair lidar-wheel_odometry \
--output docs/assets/sota_leaderboard.json --markdown standings.md
Paste standings.md into the table above.