SlugTrails

SlugTrails 1.0

SlugTrails is an egocentric floor-plan localization benchmark for large indoor buildings.
This Hugging Face package (slugtrails-1.0) is the canonical public data blob: anonymized RGB, Project Aria MPS trajectories, floor-plan–aligned ground truth, CAD-derived maps, and official evaluation / finetune splits.

manifests/ holds our official splits for the paper protocol (SHA-256 digests in RELEASE.json). You may define your own sampling if you prefer; numbers comparable to the paper should use these manifests.

What’s inside

• 3 campus buildings / 6 floors — floor-plan outline 22 089 m², circulation space 5317 m² (same figures as the paper)
• 6 Aria sessions (~30 Hz RGB); trajectory length ≈ 1.80 km (≈ 1.57 km excluding panorama spins)
• CAD floor plans with semantic classes + outline + circulation masks (meters)
• MPS open-loop and closed-loop device trajectories (Meta export)
• Factor-graph GT (gt_aligned/) in the same meter frame as the maps
• Official protocols: single-frame, multi-view sweep, walking sequence + finetune session split

Not in 1.0: raw VRS, SLAM-camera videos, MoGe depths, extra non-protocol sessions.

Quick layout

slugtrails-1.0/
├── README.md                 ← this card
├── RELEASE.json
├── CHECKSUMS.sha256
├── maps/<floor_stem>/        ← walls, doors, windows, … outline, circulation
├── manifests/                ← official eval / finetune splits
└── recordings/<recording_id>/
    ├── RECORDING.json
    ├── rgb/                  ← anonymized camera-rgb_undistorted.mp4 + timestamps
    ├── mps/{closed,open}/
    ├── gt_aligned/           ← canonical SE(3) GT @ ~30 Hz (eval lookup)
    ├── labels/
    ├── alignment/            ← coarse SE(2) prior to FG
    └── metadata/             ← sanitized Aria metadata + RGB intrinsics

Per-folder docs: maps/README.md · manifests/README.md · recordings/README.md.

Protocol recordings

Leave-out roles (finetune / paper): ID test session BE_2_0, OOD test session PS_2_0 — see manifests/finetune_session_split.json.

Coordinate frames (read this once)

Eval scoring looks up nearest gt_aligned/ pose by timestamp_ns.

Official evaluation protocols

SHA-256 digests for these files are listed in RELEASE.json (also covered by CHECKSUMS.sha256).

Privacy / anonymization

Public RGB was anonymized before release. Faces / heads were detected with SAM3, then manually inspected and blurred (elliptical Gaussian blur baked into rgb/camera-rgb_undistorted.mp4). Working label sidecars are not shipped.

We also checked that individual name tags do not appear as readable text, along with room numbers — under our QA, all such text is too blurry to read.

Although we have done our best to anonymize, if you need additional anonymization, contact the corresponding author at [email protected].

Capture is in public / circulation spaces of campus buildings.

Verify download

From a checkout of the code tools (public GitHub) with this tree as the data root:

# Example: confirm RELEASE.json + manifest digests (API may take a data root flag —
# see the public repo README for the exact verify_release invocation once published.)
sha256sum -c CHECKSUMS.sha256 | tail

Spot-check one recording: open recordings/BE_2_0/rgb/camera-rgb_undistorted.mp4, load gt_aligned/trajectory_positions.npy, confirm RECORDING.json → anonymization.status == "done".

Citation

If you use SlugTrails, please cite:

@inproceedings{Cheng2026SlugTrailsAE,
  title={SlugTrails: An Egocentric Benchmark for Floor Plan Localization in Large Buildings},
  author={Yunqian Cheng and Roberto Manduchi},
  year={2026}
}

Also cite Project Aria / MPS and any baseline methods you evaluate under their own licenses.

Contact

Yunqian Cheng ([email protected]) · Roberto Manduchi ([email protected])
UC Santa Cruz — Computer Science and Engineering

Issues / data questions: GitHub Issues on the public repo.