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, an anonymized Aria VRS (native fisheye RGB + SLAM cameras), Project Aria MPS trajectories, floor-plan–aligned ground truth, CAD-derived maps, and official evaluation / finetune splits.

Release id	slugtrails-1.0 (package 1.0.0)
License	CC BY 4.0
Code & tools	github.com/Head-inthe-Cloud/SlugTrails
Paper	Cheng & Manduchi — SlugTrails: An Egocentric Benchmark for Floor Plan Localization in Large Buildings (2026)

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)
• Anonymized VRS per session (vrs/recording.vrs): one forward RGB camera (camera-rgb, color) and two side SLAM cameras (camera-slam-left, camera-slam-right, grayscale). All three are blurred. Device calibration stays inside the VRS.
• CAD floor plans with semantic classes + outline + circulation masks (meters)
• RGB valid mask (valid_mask.png): one 1408×1408 mask for every session. White pixels are the usable undistorted RGB disk. Use it to drop the fisheye border when you filter depth on camera-rgb_undistorted.mp4.
• MPS open-loop and closed-loop device trajectories (Meta export) + semidense point clouds (mps/slam/semidense_points.csv.gz) and per-frame observations (mps/slam/semidense_observations.csv.gz) for Aria Studio / Rerun viewing
• Factor-graph GT (same map meters as the floor plans):
  • gt_aligned_device/ — device trajectory (Aria default device frame = left SLAM camera, camera-slam-left)
  • gt_aligned_camera/ — camera trajectory (upright RGB / camera-rgb, matches rgb/camera-rgb_undistorted.mp4)
• Official protocols: single-frame, multi-view sweep, walking sequence + finetune session split

Not in 1.0: the original unblurred VRS, standalone SLAM MP4s, MoGe depths, extra non-protocol sessions. Eye-tracking images inside recording.vrs are copied from the capture and are not covered by the face-blur labels.

---

Quick layout

slugtrails-1.0/
├── README.md                 ← this card
├── RELEASE.json
├── CHECKSUMS.sha256
├── valid_mask.png            ← shared 1408×1408 mask for undistorted RGB / depth
├── 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
    ├── vrs/                  ← anonymized recording.vrs + per-camera blur boxes
    ├── mps/{closed,open,slam}/  ← traj + semidense_points.csv.gz + semidense_observations.csv.gz
    ├── gt_aligned_device/    ← FG GT: device traj (left SLAM / camera-slam-left)
    ├── gt_aligned_camera/    ← FG GT: upright RGB camera traj
    ├── 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

Recording	Floor map
BE_0_0	maps/BE_0/
BE_1_0	maps/BE_1/
BE_2_0	maps/BE_2/
E2_2_0	maps/E2_2/
E2_3_0	maps/E2_3/
PS_2_0	maps/PS_2/

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)

Asset	Frame
maps/*	Floor-plan meters; origin = CAD footprint bottom-left (after unit scale). Not geographic north unless the DXF already was.
gt_aligned_device/	Same map frame. Aria device trajectory after FG alignment. On Aria Gen1 the device frame defaults to the left SLAM camera (camera-slam-left), same pose family as Meta MPS closed-loop.
gt_aligned_camera/	Same map frame. Upright RGB camera trajectory for camera-rgb / rgb/camera-rgb_undistorted.mp4 (T_world_camera = T_world_device @ T_device_camera).
mps/closed, mps/open	Meta MPS device poses (meters) in the MPS frame — not the floor-plan frame.
Timestamps	Aria device nanoseconds across RGB, MPS, labels, GT.

Score RGB→SE(2) methods against gt_aligned_camera/. Use gt_aligned_device/ when you want the Aria device / left-SLAM body (or to stay consistent with MPS pose type). Lookup is nearest GT pose by timestamp_ns.

Note: an earlier package revision left device vs camera ambiguous in the docs. This package ships both named folders and drops the old single gt_aligned/ path.

---

RGB valid mask

valid_mask.png is one grayscale image, 1408×1408, the same size as rgb/camera-rgb_undistorted.mp4. It is shared by every recording. White (255) is valid. Black (0) is the fisheye border left after undistortion (a circle of radius 900 px about the image center).

When you turn those undistorted frames into depth or a point cloud, keep a sample only where this mask is white. The border is not a measured surface. This file is not a face-blur mask, and it does not apply to the native fisheye images inside vrs/recording.vrs.

---

Official evaluation protocols

Manifest	Task	Size (this release)
direct_single_manifest.json	One walking RGB → SE(2)	1296 samples @ 1 Hz (pano / pitch filters)
direct_sweep_manifest.json	Stationary multi-view spin → one pose	2880 sweeps (180°/270°/360° tiers)
sequential_stream_manifest.json	Walking stream + relative motion	600 sequences / 26400 frame rows @ 2 Hz, seed 42
finetune_session_split.json	Train / ID / OOD session carve	roles into the three tasks

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

---

Privacy / anonymization

Public images were anonymized before release. Faces / heads were detected with SAM3, then manually inspected.

• Undistorted RGB (rgb/camera-rgb_undistorted.mp4): elliptical face blur baked into the MP4. Pair it with metadata/camera-rgb_intrinsics_undistorted.json.
• Native fisheye (vrs/recording.vrs): the same blur baked into the color RGB camera and both grayscale side SLAM cameras (camera-slam-left, camera-slam-right). Calibration is the VRS device calibration. The boxes used for that blur ship as JSON under vrs/blur/<stream>/boxes.json (face = ellipse, plate = rectangle, keyed by frame index). The eye-tracking stream is not blurred.

We also checked individual name tags on the RGB: none appear as readable text under our QA, so no further redaction was needed for names. Room numbers remain visible (building wayfinding; not personal PII).

Although we have done our best to anonymize, if you need additional anonymization, contact the corresponding author at ychen827@ucsc.edu.

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_camera/trajectory_positions.npy, confirm RECORDING.json → anonymization.status == "done".

---

Citation

If you use SlugTrails, please cite the paper: https://arxiv.org/abs/2609.19876

@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 (ychen827@ucsc.edu) · Roberto Manduchi (manduchi@ucsc.edu)
UC Santa Cruz — Computer Science and Engineering

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