Paper vs Data Contradictions

Confirmed against the files. Where a number matters for estimation, the data value governs.

Item Papers Data / reality
Static-IMU duration 5 h (body) / 4 h (table, README) measured ~3.67 h
Fiducial markers 108 (body) vs 125 (tables) the public field metadata defines 108 transition-field tags on ids 0..107,
plus one extra standalone 0.18 m tag 123;
ids 108..122 are unused in tag_info.yaml.
The field is reconstructable from tag_info.yaml + the board geometry (gen_tag_board_calib.m) + the marker pipeline (What Is Not in the Download),
and the geometry carries a transcription typo (Known Defects)
GT attitude accuracy implied cm / sub-degree truth position cm;
attitude gravity-inconsistent 0.9–55° median (≤~77° p90), mission-dependent;
Wahba uses raw px4_mag (intrinsic not applied);
KLU-site GT not pipeline-reproducible (Source of Truth)
LSM mag intrinsic (calibration shipped per site) the lsm_mag_cal_intr blocks are swapped between the two site YAMLs (Calibration); PX4 not swapped
IMU count “four” (body) vs “three” (table) 3 usable streams ship; the 4th (autopilot BMI055) is deactivated
GT rate “8 or 50 Hz” (tables) files are 8 Hz and 100 Hz-grid (“80hz”); no 50 Hz file
RTK baseline 1.2 m 1.16 m (sensor_calibration.yaml); measured 1.15–1.19 m
Companion boards Raspberry Pi 4 ×2 (Odroid XU4 belongs to the CNS Flight Stack demonstrator [1], not this rig)
Mocap rate 300 Hz (Table 1, §3.2 prose, appendix Tables 7–9)
vs 360 Hz (Figure 3 data-rate diagram) — the paper is self-inconsistent		 ~334 Hz median instantaneous (per-run 333.5–334.9 Hz; pooled 333.7 Hz over all six mocap runs), between the paper’s own two figures

The two distributed trajectory summaries also disagree with each other: the download’s dataset_start_here.txt (website table) and the paper’s Table V give different per-run distance/velocity/height figures — e.g.

Neither table is the trajectory; integrate the per-run ground_truth_* for an actual path length.

References

[1]
M. Scheiber et al., “CNS flight stack for reproducible, customizable, and fully autonomous applications,” IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 11283–11290, 2022, doi: 10.1109/LRA.2022.3196117.