KITTI-360 camera-LiDAR: multi-window check (development protocol)

Status, 2026-11-01: run once; verdict not carried forward. All six criteria held, but one of the 42 segments could not run (an archive data gap), and the protocol required all 42. This is development, not a claim. KITTI-360 holds no unread camera-LiDAR data here (the exposure audit), so nothing on this page is a SOTA or held-out result. The 42 segments of 0004–0007, 0009 and 0010 have not been read by this development work. They serve as its held-out check, run once under the rule below. The rule and the criteria are committed before that run, and neither changes after it.

Why

On spent segment 0002-s04 the check flags KITTI-360's calibration. It reads yaw −0.60°, with a reported std of 0.042°. Both halves of that minute read the same bias, so a block jackknife within one window cannot see it (cause). Disjoint windows of one recording see different scenes. Their spread measures a bias a single window shares across its blocks, and their median is robust to one biased window. Raising the 0.5° floor would hide the flag too, but it is a threshold change and is not considered.

Rule

tools/kitti360_camera_lidar_multiwindow.py combines the per-window results of the unchanged default check. A recording of K disjoint 60 s windows gets one verdict per rotation axis, over the windows whose check judges that axis:

• Too few windows. Fewer than max(2, ceil(K / 2)) judging windows leaves the axis unchecked.
• Center. m is the median of the windows' candidate errors (candidate minus estimate).
• Spread. s = 1.4826 · MAD of those errors, w is the median reported std, and σ = max(s, w). The median's std is 1.2533 · σ / sqrt(n).
• Tolerance. max(3 · that std, 0.5°), the check's sigma_k and rotation floor.
• Status. Pass when |m| is within the tolerance, fail beyond twice it, otherwise warn.

The recording's verdict is its worst axis, or inconclusive when no axis is judged. Turned runs use the same injection in every window. A trial counts when the multi-window baseline judges the turned axis. It is flagged when the turned run's multi-window status is warn or fail, as in v2. Accuracy uses the median of the judging windows' estimate errors against the KITTI-360 calibration.

Recordings and criteria

The exposed segments are combined two ways:

• Sequence level. Each sequence's segments, in order: 6 recordings, with K from 3 to 12.
• Triple level. Disjoint consecutive triples within a sequence: 13 recordings (0004: 3, 0005: 2, 0006: 2, 0007: 1, 0009: 4, 0010: 1).

The unchanged single-window check is reported on the same 42 segments for comparison. The rule is carried forward only if every one of the 42 segments runs and all six criteria hold:

Criterion	Requirement
Vendor calibration	no sequence- or triple-level recording flags it
Baseline accuracy	every judged triple-level baseline axis within 1.0°
Turned-run accuracy	every judged triple-level turned-run axis within 1.0°
3° turns	at most 0.15 of counted triple-level trials missed, none judged right
1° turns	at most 0.25 of counted triple-level trials missed
Coverage	at least 0.5 of triple-level baseline axes judged

The accuracy and detection bounds are the KITTI-360 development gates (gates). The vendor requirement is v2's. If carried forward, the rule becomes an opt-in calibrex check option for long recordings. Otherwise it is recorded as not carried forward.

Design data

The rule was written while looking at the 22 spent segments only. On them, the single-window check flags the vendor calibration once (0002-s04), as the v2 scorer does. A unit test pins that parity. The 7 spent triples and 2 spent sequences flag it never. Their worst judged errors are 0.55° (baseline) and 0.60° (turned). They miss 2 of 19 counted 3° turns and 0 of 19 1° turns, and judge 19 of 21 baseline axes. These are design data, not evidence.

Result (run once, 2026-11-01)

The batch ran all 42 segments from the protocol commit 41ee8249. The rule was applied once to the published receipts. 41 of 42 ran. 0009-s07 failed at fetch: the official OXTS archive has no packet for frame 8849. The batch requires every member, so the attempt was not retried, since a retry would fail the same way. The gap is visible in archive metadata. The segment rule checked only image and scan continuity, which is a lesson for any future segment rule. Without s07, 0009's third triple is s06, s08 and s09, giving 12 triples instead of 13.

41 scored segments	Single window (shipped check)	Triples	Sequences
Recordings	41	12	6
Vendor calibration flagged	0	0	0
Worst judged baseline axis	0.40°	0.40°	0.40°
Worst judged turned-run axis	4.75°	0.65°	0.65°
Judged baseline axes	104 of 123	33 of 36	15 of 18
3° turns missed	5 of 104	1 of 33	1 of 15
1° turns missed (judged right)	4 of 104 (2)	0 of 33 (0)	0 of 15 (0)

Frozen verdict: not carried forward. Every criterion holds at the triple and sequence levels, but the protocol required all 42 segments to run. It stays unchanged.

After the verdict (post hoc, changes nothing)

• No vendor flag to remove. The single-window check also flags none of these 41 segments. So the run does not show the rule removing a vendor flag. That benefit was seen only on the design data (0002-s04).
• Large errors removed. The rule removes the single-window check's largest errors. In 0007-s02 under a 3° roll turn, a single window judges roll 4.75° from the reference with a reading of the wrong sign (−1.75°). The trial is not counted, because the baseline does not judge roll there. Two 1° turns are judged right by a single window (0006-s05 roll, 0009-s01 yaw). No triple or sequence does either.
• A constant offset from the vendor calibration. Every triple and sequence that judges roll puts the estimate +0.35° to +0.40° from the vendor calibration in roll, and pitch 0° to −0.20°. That matches 0000 and 0002. A difference this constant across the rig is a property of the rig or of the method, not of a scene. Which one this data cannot tell.

All KITTI-360 sequences are now spent. Shipping the rule as an opt-in check option would be a product decision on development numbers, as with the disagreement warning. It would not be a validated rule.