What Is Not in the Download

• Per-frame fiducial detections — the per-frame tag-in-camera poses (fiducial_marker.csv, written by extract_data.m from the ROS topic /marker_detections) and tags.csv do not ship; marker poses also survive inside the derived transition_gt_in_{gps,moap}.csv. (Figure 31 advertises tags.csv as a per-frame detection product; the present transition-run mocap_tag_board.csv is a different product — the 6-DoF mocap pose of the static tag-board rigid body, Motion Capture — not a substitute for it.)
  • When the maintainers extract fiducial_marker.csv, it is a dynamic wide CSV with t followed by ex_<id>, p_x_<id>, p_y_<id>, p_z_<id>, q_w_<id>, q_x_<id>, q_y_<id>, q_z_<id> for each detected tag id, with absent-tag fields left at zero on rows where that tag is not present.
  • But the tag field is reconstructable — and this is the load-bearing thing the earlier framing missed: tag_info.yaml (108 transition-field tag ids on 0..107, plus standalone tag 123, with ids 108..122 unused), the exact physical board geometry (fiducial_marker/tag_calibration/gen_tag_board_calib.m — tag layouts for all 8 boards), and the complete marker→trajectory pipeline (tag_calibration/, tag_evaluation/tags_to_trajectory.m, ArUco tag_reading/read_aruco*.m, post_scripts/export_tag_data.py) are all in insane_dataset_tools [1].
  • In that board geometry, the local board frame origin is the bottom-left corner, and the laid-out tags imply a 2.0 m × 0.6 m board rectangle with about 4 cm outer margins around the outermost tag edges.
  • The field is two board classes plus a standalone tag: five Type-1 boards (tag ids 0–14, 15–29, 30–44, 45–59, 60–74 — 15 tags each) and three Type-2 boards (75–85, 86–96, 97–107 — 11 tags each), plus the standalone 0.18 m tag 123. Marker edge size is per-tag and heterogeneous in tag_info.yaml (0.10 / 0.15 / 0.20 / 0.30 / 0.50 m): each Type-2 board carries a single large 0.50 m anchor tag (ids 80, 91, 102; 102 is the pipeline main_tag_id on board 8, the mocap-tracked tag_board_8), while the rest run 0.10–0.30 m — so single-tag PnP must apply the correct per-tag size for metric scale. The three large-element boards are the ones the paper reports detectable up to ~10 m altitude for the outdoor phase; the smaller-element boards serve the close-range indoor/transition phase.
  • Each board also carries small markers along its edges; their stated purpose (Figure 10 caption) is higher-accuracy board-to-board calibration through the wide lever arm at the board extremities. The shipped maintainer calibration and the pooled fiducial map both weight all tags uniformly and do not single these out.
  • So the maintainers’ transition marker bridge can be regenerated from the raw nav-cam PNGs + this geometry (detected ArUco DICT_7X7_250, the field matches the maintainer design to ~1% / a few mm).
  • The board geometry carries a transcription typo that corrupts a design-rigidified map if used verbatim (Known Defects), and one physical build irregularity the maintainer file itself flags: board 1’s tag id 3 is not rotated 45° like the Type-1 template and boards 2–5 (gen_tag_board_calib.m r_world_yaw carries 0 there, with an IMPORTANT comment), so a template-based board-1 map mis-orients that tag by 45°.
  • The public metadata supports the paper’s 108-field-marker count; the extra standalone tag 123 is outside that transition field.
• No camera↔︎LRF calibration product (no static-wall/IR-indicator recording; no calibration script in the tooling).
• No stitched full-run transition truth lane (Source of Truth).
• No anchor-position table for UWB — but the three ground anchors are board-mounted (boards 1/2/7) and recoverable in the board frame, and are observable by trilateration from the moving vehicle (Ultra-Wideband Ranging); the range stream is not unusable for absolute localization. No device-model strings beyond LSM9DS1/Decawave.
• No precomputed Kalibr results for klu1/klu2/stereo (only the bags + kalibr.sh); the cam↔︎IMU timeshift ships only in the mars nav-cam raw bundle.
• Physical RPM (Wachendorff PT99) is not in px4_rpm.csv; only the normalized command is.

References

[1]

aau-cns, INSANE Dataset Tools. GitHub. Accessed: Jun. 09, 2026. [Online]. Available: https://github.com/aau-cns/insane_dataset_tools