Skip to content

Short-baseline neighbors restore image-only motion support

Acquisition scope frozen 550055c0; exact byte ranges frozen ce811402 before payload reads. Audit freeze 11f93d00, input preprocessing reconciliation 79ce9e80. See protocol, metadata index, acquisition manifest, acquisition schema, final audit plan, all36 pair results, result schema and verification.

Image-only directional-support screening passes 28/36 short-baseline pairs. Intervals are .103129–.103757 seconds, versus3.728–11.590 seconds for the earlier sparse triplets (1/12 passed). These have different pair denominators, so this is an input-feasibility finding, not a paired accuracy improvement or a recovered camera/LiDAR calibration. Metric translation scale remains unvalidated.

Already-spent recording Adjacent pairs passing All pairs retained
2011_09_26_0005 6 6
2011_09_26_0009 6 6
2011_09_26_0011 4 6
2011_09_26_0013 6 6
2011_09_26_0015 6 6
2011_09_28_0002 0 6

Median cheirality-supported feature count is340.5. The two0011 failures correspond to frame173→174 and174→175:120/116 inliers but only5/4 cells per image, below the six-cell rule. All six0002 pairs fail through count or spatial coverage. These failures remain; thresholds were not lowered and no pair was replaced.

Acquisition and input identity

Only72 PNG images (36 stereo neighbors around18 existing centers) were fetched, from the six declared spent archives. Archive size/ETag, exact ranges, compressed sizes, CRC32 and decoded sizes were pinned; If-Match and extraction checks held. Total requested payload ranges:59,606,871 bytes (56.85MiB); decoded image payload: 59,303,247 bytes. All per-drive32MiB and aggregate128MiB range limits hold. Metadata/HEAD and bounded transport retries are separately disclosed overhead. No full archive, LiDAR scan, new recording or fresh holdout was downloaded. All images decode, SHA256 values verify, and paths are retained under /workspace/work/temporal-neighbors-20261030. Right images are prepared for future work but not used by this monocular support audit.

The first audit aborted before producing a case: original-center grayscale differed from archived images, as the builder applies histogram equalization. The original audit plan and traceback are kept. The reconciliation note pins the builder's existing operation. Applying equalizeHist to every raw center reproduces the archived center pixel-for-pixel (18/18). The same operation is applied to all new left neighbors, with derived paths/hashes retained. No matching, RANSAC or support threshold changes occurred. This restores consistent photometric preprocessing; it is not an outcome-selected image enhancement.

Every short-baseline pair

Drive Frame pair Cheirality inliers Cells Pass
2011_09_26_0005 37→38 336 [11, 12] True
2011_09_26_0005 38→39 394 [11, 12] True
2011_09_26_0005 75→76 724 [11, 11] True
2011_09_26_0005 76→77 624 [12, 13] True
2011_09_26_0005 113→114 686 [11, 11] True
2011_09_26_0005 114→115 608 [11, 11] True
2011_09_26_0009 110→111 468 [15, 15] True
2011_09_26_0009 111→112 487 [12, 14] True
2011_09_26_0009 222→223 323 [9, 11] True
2011_09_26_0009 223→224 338 [9, 10] True
2011_09_26_0009 333→334 492 [10, 11] True
2011_09_26_0009 334→335 462 [11, 13] True
2011_09_26_0011 57→58 520 [12, 12] True
2011_09_26_0011 58→59 440 [12, 12] True
2011_09_26_0011 115→116 80 [7, 7] True
2011_09_26_0011 116→117 156 [7, 7] True
2011_09_26_0011 173→174 120 [5, 5] False
2011_09_26_0011 174→175 116 [4, 4] False
2011_09_26_0013 34→35 389 [10, 10] True
2011_09_26_0013 35→36 401 [10, 10] True
2011_09_26_0013 70→71 405 [13, 12] True
2011_09_26_0013 71→72 412 [11, 13] True
2011_09_26_0013 106→107 343 [13, 13] True
2011_09_26_0013 107→108 320 [12, 13] True
2011_09_26_0015 73→74 324 [10, 10] True
2011_09_26_0015 74→75 290 [11, 11] True
2011_09_26_0015 147→148 175 [8, 8] True
2011_09_26_0015 148→149 245 [7, 7] True
2011_09_26_0015 221→222 530 [10, 11] True
2011_09_26_0015 222→223 456 [10, 10] True
2011_09_28_0002 92→93 37 [5, 5] False
2011_09_28_0002 93→94 0 [0, 0] False
2011_09_28_0002 186→187 60 [5, 5] False
2011_09_28_0002 187→188 25 [4, 4] False
2011_09_28_0002 280→281 2 [2, 2] False
2011_09_28_0002 281→282 0 [0, 0] False

Twenty-seven focused tests pass (24 archive/range/extraction tests, three image-audit/camera tests, latter rerun after preprocessing reconciliation). Repository Ruff, strict documentation build, schemas, source/input/derived-image hashes, byte caps,72-image completion,18-center pixel parity and36-case completion were checked. No production/default changes or extrinsic fits occurred.

Next validate metric stereo visual odometry on these short intervals, using known camera baseline for scale and an independently specified controlled-motion test. It must not consume reference camera/LiDAR extrinsics, and must retain all eight failed support pairs rather than silently evaluating only28. Depth and motion inferred from the same stereo stream have correlated errors; temporal agreement would be an extra measurement check, not independent ground truth. Freeze motion estimator, failure accounting and scale/time/frame conventions before execution. Only then decide whether a temporal surface-consistency experiment is justified.