The Outdoor-to-Indoor Transition Handoff

The transition handoff (transition_1..3) is the dataset’s hardest truth structure. There is no single stitched full-run truth lane. Three regimes:

1. Outdoor: ground_truth_8hz/80hz.csv (PX4-IMU, ENU) — covers the early outdoor segment; on transition_3 the 80hz lane numerically continues through the gap but is unreliable there (below).
2. Building-approach: marker-based pose, surviving only inside the derived transition_gt_in_{gps,moap}.csv helpers.
3. Indoor: mocap_vehicle_data.csv (rigid-body, mocap world) — covers the late segment only.

Coverage differs structurally between runs:

• transition_1: outdoor GT then a 4.97 s gap then mocap (sequential, no overlap).
• transition_2: same shape, 5.50 s gap.
• transition_3: the outdoor 8hz lane is split by a 77.75 s gap (≈[113.8, 191.5] s rel. IMU start) and the contiguous mocap_vehicle segment (≈42.8 s, [124.9, 167.7] s) falls entirely inside that GT8 hole.
  • So there is no co-temporal overlap between ground_truth_8hz and mocap_vehicle on any transition run (an earlier “overlap ~70 s” reading was wrong; what spans the whole run is the static mocap_tag_board lane, not the vehicle pose).
  • The ground_truth_80hz lane does span this gap (its own largest interior gap is 0.05 s), but its in-gap trajectory is unreliable: strict-gap net displacement 8.0 m vs 2.6–3.4 m from three independent estimates (stereo VO, rs_odom, the fiducial marker bridge), a ~2.4× overshoot — do not use 80hz as truth inside the transition_3 gap (verified by the camera-feasibility probe).
  • transition_3 also lacks time_info.yaml and a <run>_sensors/README.txt.
  • Its shipped GT attitude is ~16° gravity-inconsistent like the other KLU runs (a “26°” sometimes quoted is a regeneration artifact, not the shipped product).

All three transition runs nevertheless have real raw dual-RTK overlap with mocap_vehicle even though none has co-temporal overlap between ground_truth_8hz and mocap_vehicle.

• Quantitatively, with the cleaned mocap_vehicle lane used by dir-5: transition_1 has 39.09 s / 313 dual-RTK paired rows inside the overlap window, transition_2 has 35.48 s / 282 rows, and transition_3 has 42.77 s / 339 rows.
• On all three runs those overlap rows carry rtk_status == -1 on both receivers, even though the rows still contain positions and gps_fix_type can be 2 or 3, so the overlap is real geometry on disk but diagnostic-only under the published RTK status semantics.

To evaluate across a full transition trajectory you must stitch the three lanes yourself, accounting for their different frames (ENU vs mocap-world), the PX4-IMU-vs-rigid-body reference-point difference, and the gaps. The _in_gps and _in_moap helpers are two different aligned frames (tens of metres apart) and are not primary scoring truth.

Gap and detection-window geometry (windows in seconds rel. IMU start; the truth gap is outdoor-GT8-end → mocap-start):

Run	Outdoor GT8	mocap_vehicle	Truth gap	nav-cam >=4 tags, raw full-run scan	nav-cam >=4 tags, undistorted full-run scan
transition_1	[0.8, 168.2]	[173.2, 212.2]	4.97 s	[150.5, 196.6]	[151.2, 182.3]
transition_2	[0.6, 174.9]	[180.4, 221.0]	5.50 s	[156.1, 202.9]	[156.9, 185.1]
transition_3	[0.3, 113.8] ∪ [191.5, 194.9]	[124.9, 167.7]	11.1 s	[93.7, 170.6]	[97.5, 170.6]

The nav-cam >=4-tag window overhangs the gap on both sides on every run under both the raw full-run scan and the undistorted full-run scan, so the fiducial field bridges outdoor and indoor in time. The dir-6 gate and the dir-7 fiducial front-end use the undistorted detector path. (transition_3’s 11.1 s truth gap is the leading slice of the 77.75 s GT8 hole, before mocap begins.)

In-gap witness coverage (counts are samples strictly inside the truth gap, per run t1/t2/t3) — the gap is narrow and several non-GT sensors cover it continuously, so it is reconstructable:

• nav-cam fiducials (downward camera, ArUco DICT_7X7_250):
  • ≥4 tags on 99 / 110 / 222 gap frames; the tags form a single connected co-visibility component, but no tag is seen on both immediate handover flanks — so the field yields absolute, anchored in-gap localization (tied to the static mocap_tag_board, Motion Capture), not a direct outdoor↔︎indoor tag bridge.
  • A full-run scan later does find transition_3 shared outdoor/indoor tags, but only far deeper into the indoor side, outside the local pre/post-gap flank pair that matters for a direct tag-only bridge: the raw-image scan finds 47, 49, 51, and 53, while the undistorted detector path used by dir-6/dir-7 finds 51 and 53.
  • Detection coverage is not a usable PnP fix everywhere: on transition_1 all 99 gap frames see ≥4 tags yet planar-PnP degeneracy leaves only ~21 with reprojection < 3 px.
• LRF (lrf_range): 100% valid through every gap, ~30 Hz (149 / 164 / 332 samples), range 0.86–1.47 m — the tightest in-gap vertical witness (~37 mm agreement).
• baro (px4_baro): 100% temporal coverage, ~18 Hz delivered (92 / 102 / 203 samples); independent vertical physics but coarse (5–10 cm random noise) with a dynamic-pressure bias during the powered descent.
• UWB (uwb_range): 27 / 30 / 50 in-gap ranging rows (≥1 valid anchor on 27 / 28 / 50, up to 3 anchors), 3–8.5 m — horizontal information, not an absolute 2-D fix.
• rs_odom: finite through the gap on transition_1/transition_3 (100%) but all-NaN on transition_2.