Camera-only metric motion passes controlled geometry and most internal checks¶
Frozen 271b9e75, evaluated2026-10-30. See
protocol,
plan,
18 controlled cases,
all36 real pairs,
schema and
verification.
All18 known-motion controls pass the frozen<.2deg/<.05m gate. Maximum rotation error .04315deg, maximum translation error .005561m. These tests establish the metric scale/frame convention and robustness of PnP on generated correspondences with depth/pixel noise and outliers; they do not validate the full real image matching/stereo pipeline against external motion truth.
After controls passed, all36 real pairs were evaluated. 34/36 pass the declared internal support/cycle checks. Metric scale comes from camera stereo calibration; no LiDAR points, target camera/LiDAR transform or OXTS camera-motion conversion enters estimation. Stereo estimates and cycle checks share measurements, so a passed check is not independent real-motion accuracy or adoption evidence.
| Recording | Internal support/cycle passes | All pairs retained |
|---|---|---|
| 0005 | 6 | 6 |
| 0009 | 6 | 6 |
| 0011 | 6 | 6 |
| 0013 | 4 | 6 |
| 0015 | 6 | 6 |
| 0002 | 6 | 6 |
Failed0013 pairs106→107 and107→108 have inlier ratios below50% in both directions: 298/644 and261/610;267/599 and280/574 respectively. Spatial support and cycle alone would pass; the ratio gate is retained. No threshold or frame is changed. All eight previous essential-matrix screen failures are evaluated; stereo-PnP uses a different measurement model and need not share its cheirality limitations, especially for little translation. This is not a paired accuracy claim.
Cycle rotation residuals span .000081–.08383deg, median .03353deg; translation .0000295–.06907m, median .01547m. All36 cycles pass the numeric cycle bounds, including the two low-ratio failures. A mutually consistent wrong depth scale or motion bias can remain invisible to forward/reverse agreement.
Input and implementation checks¶
All54 unique stereo-frame depth arrays were computed using pinned grayscale SGBM, re-rectification and restoration. All18 central arrays exactly reproduce the archived SGBM outputs. Depth paths/hashes are retained in each result and verified; arrays reside under /workspace/work/stereo-metric-motion-20261030. Both PnP directions, inlier identities, counts, cells and target-from-source transforms remain in the report. The image matching is on histogram-equalized images, stereo on raw grayscale.
Three focused tests, repository Ruff and strict documentation build pass. Tests verify metric translation direction, insufficient/missing-depth failure handling and blank-image zero support. Integration checks validate schema, all source/input hashes,54 depth hashes,18 controlled outcomes, center parity and36-case completion. No core/default changes, new downloads, extrinsic fit, ICP rerun or SOTA claim.
Next measurable step¶
Freeze a temporal surface-consistency diagnosis, beginning with independently specified controlled camera motion and visibility/occlusion handling. On real frames use only the estimated camera-to-camera transforms, preserving both failed motion pairs and out-of-view/occluded/unsupported pixels explicitly. Score surface quality with reference/LiDAR only AFTER selection; never use the target extrinsic to produce a temporal mask. Independent metric real-motion truth remains absent, and shared stereo errors must be disclosed. Retain every original surface failure; temporal self-consistency is an extra check, not proof of correct geometry or calibration recovery. No real-data temporal filter is run in this stage.