tjelinek/slaminthewild

SLAM in the Wild

Contents

proxies/            three chained 720p H.264 proxies of the ride (25 fps, ~2.5 GB total)
                     decode ~20x faster than the source 4K footage and carry
                     the same CreateDate, so extracted frame timestamps match
ground_truth/        1 Hz smartwatch GPX track for the full activity
calib/                camera intrinsics used for the proxies (UNCALIBRATED
                      FOV guess, see Limitations)
results/
  orb-slam3/          f_<tag>.txt, kf_<tag>.txt (EuRoC-format trajectory,
                      nanosecond timestamps), result.json
  mast3r-slam/        traj_tum_ns.txt, points_2m.npy
  droid-slam/         traj_tum_ns.txt, points_2m.npy
  lingbot-map/        traj_tum_ns.txt, points_2m.npy
recordings/           one self-contained Rerun (.rrd) recording per method:
                      3D trajectory vs. GPS, input frames, OpenStreetMap
                      overlay, error/speed plots. Open at
                      https://app.rerun.io/version/<pin-me>/?url=<file-url>
                      or `pip install rerun-sdk && rerun <file>`

Raw 4K source video (~59 GB across three files) is not included here; see Raw footage below.

Benchmark results

Sim(3)-aligned (scale-included) against the 1 Hz watch GPX. Drift = ATE / path length.

Method	Coverage	Path length	Drift
ORB-SLAM3	373 / 22,488 frames, one 27 s map	30 m	16 %
MASt3R-SLAM	742 keyframes, one 5-min segment	391 m	15.8 %
DROID-SLAM	every frame, same 5-min segment	573 m	6.0 % (best)
LingBot-Map	full 30 min, 22,488 poses, no loop closure	3.9 km	9.5 % (~370 m absolute wander)

ORB-SLAM3 and MASt3R-SLAM lose tracking on this footage and re-initialize repeatedly (ORB-SLAM3: ~440 times in 30 min); their trajectories cover only the largest surviving map, not the full ride. DROID-SLAM and LingBot-Map segment the run instead of tracking continuously end to end. None of these numbers should be read as "the" accuracy of a method in general — they are what each one did on this specific, adversarial clip.

Limitations

• calib/guess_lrv.json is an uncalibrated FOV guess (fx wrong by an unknown few percent), not a checkerboard calibration. Treat absolute scale from any method here as approximate.
• The GPX is 1 Hz smart-recording from a wrist GPS; it is a global reference for scale and long-range error, not a metric reference for local drift.
• This footage shows other orienteering competitors' faces. Do not use it for anything beyond SLAM/trajectory evaluation research.

Raw footage

The source 4K clips (~59 GB) are available on request / via a future torrent link — TODO once published. Contact: tomasjelinek96@gmail.com.

License

CC BY 4.0. Attribute as "SLAM in the Wild (https://github.com/tjelinek/slaminthewild)".

Citation

TODO once this has a DOI / paper.