OverTheReality/OverMaps360_50

🌐 OverMaps360-50 Dataset

📄 Project page: ovr-platform.github.io/OVR-MAPS/360.html

OverMaps360-50 is the 360° companion of OverMaps-1K. Where OverMaps-1K is built from smartphone photos, every scene here is a continuous walk-through recorded with an Insta360 X5 dual-fisheye camera, captured simultaneously with a smartphone running the OverTheReality app. Each scene ships the complete chain from raw sensor data to a simulation-ready, physics-certified environment: censored 360° video, fisheye frames, 1 kHz IMU, GPS, ARKit smartphone poses, a metric COLMAP reconstruction with 5 perspective views per fisheye frame, a dense point cloud, a textured mesh with a closed floor, a 3D Gaussian Splatting model, three mesh LODs, and a MuJoCo / Isaac Sim bundle with its certification report.

Key Statistics

Scenes	50 real-world walk-throughs (6 countries: Thailand, Latvia, United States, Indonesia, Spain, Mexico)
Fisheye frames	118,816 (3840×3840, two lenses), median 2,404 per scene
Perspective views	594,080 pinhole views (1600×1600, 90° FOV, 5 per fisheye frame), median 12,020 per scene
Smartphone photos	5,591 ARKit-tracked photos with metric poses, median 72 per scene
IMU	29.8 M gyroscope + accelerometer samples at 1 kHz, synchronised to the video
Walked distance	16.3 km in total, median 316 m per scene
Certified walkable area	180,000 m² in total, median 3,548 m² per scene
3D Gaussian Splatting	250 M Gaussians (5 M per scene)
Dense points	381 M dense reconstruction points, median 7.4 M per scene
Metric scale	every asset of a scene shares one metric frame, scaled from the ARKit smartphone trajectory
Volume	1.13 TB, median 22 GB per scene

---

⚠️ Disclaimer

This dataset is a free research-only sample of 50 scenes extracted from a larger collection of 360° walk-throughs. The full collection is available under a commercial license. For access or licensing inquiries, please contact data@ovr.ai.

---

📱 How it was acquired

1. The capture setup

Each scene is recorded by one operator walking through the environment with:

• an Insta360 X5 on a pole above the head, recording dual-fisheye 360° video at 3840×3840 per lens (~5.6 K equivalent), with the camera IMU (gyroscope + accelerometer, 1 kHz) embedded in the recording;
• a smartphone running the OverTheReality app, which tracks the walk with ARKit visual-inertial odometry and takes tracked photos along the way. The phone provides the metric scale, GPS at 1 Hz, compass, and optional audio clips.

The walk is designed to cover the scene from the ground: streets, squares, parks, temples, museums, malls, campuses. Camera height is about 2.2 m.

2. The "map2earn" incentive

As for OverMaps-1K, contributors are compensated only after their mapping passes the validation pipeline (the map2earn model), which enforces protocol compliance and the choice of geometrically interesting scenes.

3. Processing pipeline

Every scene went through the same automated pipeline:

1. Frame extraction: fisheye frames are extracted from the 360° recording at a fixed rate (~2 fps), one JPG per lens.
2. Privacy masking: people and license plates are segmented on both lenses and on the smartphone photos. The masks are shipped and were applied to everything published: fisheye frames, perspective views, smartphone photos and the 360° video (360recording_censored.insv). Masked pixels are blacked out (no inpainting).
3. Matching and Structure from Motion: learned image matching with sequential, cross-lens and place-recognition loop-closure pairs, then a rig-aware SfM with a fisheye camera model per lens and a fixed two-lens rig; the smartphone photos are registered jointly in the same reconstruction.
4. Metric alignment: the reconstruction is scaled and gravity-aligned to the ARKit trajectory of the smartphone (validated against ground-truth markers).
5. Perspective views: each fisheye frame is re-projected into 5 pinhole views (1600×1600, f = 800 px, 90° FOV: front, left, right, up, down), with the masks carried over. This perspective model is what the dense reconstruction and the 3DGS training consume.
6. Dense reconstruction and mesh: multi-view stereo on the perspective views, then a textured mesh whose floor is closed with a monocular geometry prior, so robots and agents never fall through gaps between the walked corridor and the walls.
7. 3D Gaussian Splatting: trained on the perspective views, 5 M Gaussians per scene, initialised from the dense cloud.
8. Simulation bundle and certification: the mesh is converted into a MuJoCo scene (height-field ground + convex collision parts), a USD scene for Isaac Sim, a navmesh and three mesh LODs; each scene is then gated by 8 physics checks in MuJoCo and a drop test in Isaac Sim with ovr-maps-360-simkit (see Certification). All 50 scenes are tier T1a.
9. Automated annotation: a Vision-Language Model analyses the capture to produce a caption, scene type, lighting, weather and crowd density (in the manifest).

---

📂 Dataset Structure

The dataset is one folder per scene, named by the scene UUID. Every scene is self-contained: download a single folder and you have everything for that scene.

OverMaps360-50/
├── README.md
├── LICENSE
├── dataset_manifest.csv / .parquet          # one row per scene: paths + annotations
└── <uuid>/
    ├── 360recording_censored.insv           # original Insta360 X5 recording, people/plates blacked out
    ├── insta360_imu.csv                     # 1 kHz gyro + accel from the recording, t relative to video start
    ├── backgroundlocationData.txt           # smartphone GPS at 1 Hz: epoch_s, lat, lon, accuracy_m
    ├── fisheye/lens0.tar, lens1.tar         # fisheye frames lensN_XXXXXX.jpg, 3840×3840, censored
    ├── masks.tar                            # privacy masks: lensN_XXXXXX.png (fisheye), <seq>_NNNN.png (smartphone)
    ├── colmap_perspective/
    │   ├── sparse/0/                        # COLMAP binary model (cameras, images, points3D, rigs, frames)
    │   ├── images.tar                       # 5 perspective views per fisheye frame + smartphone photos, JPG
    │   └── masks.tar                        # one 8-bit PNG per view (0 = masked)
    ├── smartphone_raw.tar                   # smartphone photos as shot (1920×1080), censored
    ├── smartphone_poses_<seq>.csv           # ARKit pose, intrinsics, GPS, compass, IMU per smartphone photo
    ├── smartphone_audio/*.wav               # audio clips recorded by the app (45 of 50 scenes)
    ├── training_cameras.json                # cameras used for the 3DGS training (800×800)
    ├── gaussian_splatting.ply               # 3DGS model, 5 M Gaussians, standard 3DGS PLY (f_dc + f_rest)
    ├── gaussian_splatting.splat             # same model in .splat format (32 bytes per Gaussian)
    ├── dense.ply                            # dense point cloud
    ├── mesh/model.glb, model.usdz, *.png    # textured mesh with closed floor (glTF and USDZ)
    ├── lod/<id8>_{high,mid,low}.ply + .json # mesh LODs (p95 deviation ≤ 1 / 3 / 10 cm), pruned of debris
    ├── simulation/<id8>.sre/                # simulation bundle (see below)
    ├── simulation/certification/            # MuJoCo gate + Isaac Sim drop test reports
    └── datasheets/<uuid>.json               # per-scene datasheet (tier, stages, QA)

Image collections (thousands of files per scene) are shipped as uncompressed .tar archives that extract in place, next to the archive, into the folder the COLMAP model and the manifest expect:

cd OverMaps360-50/<uuid>
tar -xf colmap_perspective/images.tar -C colmap_perspective   # -> colmap_perspective/images/
tar -xf colmap_perspective/masks.tar  -C colmap_perspective   # -> colmap_perspective/masks/
tar -xf fisheye/lens0.tar -C fisheye && tar -xf fisheye/lens1.tar -C fisheye
tar -xf masks.tar && tar -xf smartphone_raw.tar

<id8> is the first 8 characters of the scene UUID; <seq> is the smartphone capture sequence id.

Coordinate conventions

• Source frame (colmap_perspective/sparse/0, dense.ply, mesh/, gaussian_splatting.*, training_cameras.json, lod/): one metric frame per scene, gravity along −Y (up = −Y), floor at y ≈ 0, cameras at y ≈ −2.2 m. Cameras follow the COLMAP convention (images.bin stores world-to-camera rotations as quaternions qw qx qy qz and translations).
• Simulation frame (simulation/<id8>.sre/): Z up, floor at z = 0, metres. The rigid transform from the source frame is stored in frame/transform.json (sim_from_source, 4×4).
• Smartphone poses (smartphone_poses_<seq>.csv) are in the ARKit frame of the phone session (Y up, metric), not in the scene frame. The registered smartphone photos in colmap_perspective/sparse/0 give the same shots in the scene frame.

Perspective views and naming

Each fisheye frame is re-projected into 5 pinhole views sharing one pose id: <pose>_perspective_<face>.jpg with face ∈ {00000000, 00000001, 00000002} for the three horizontal directions, 00000004 up and 00000005 down. All are SIMPLE_PINHOLE 1600×1600 with f = 800 px (90° FOV). Smartphone photos in the same model are <seq>_NNNN.jpg, 1080×1080 (rotated and square-padded), SIMPLE_RADIAL. The COLMAP model contains the rig and frame tables (rigs.bin, frames.bin) that group the 5 views of one pose.

Simulation bundle (simulation/<id8>.sre/)

Path	Content
scene.xml	MuJoCo scene: height-field ground, convex collision parts, walk-crop
scene.usda	the same scene as USD for Isaac Sim
collision/surface.ply	the collision surface (mesh cropped to the walked corridor, in the simulation frame)
collision/part_NNN.obj	convex collision parts
collision/ground_hfield.png, navmesh_grid.npy, navmesh_ground_z.npy	ground height field and walkable-cell navmesh (0.25 m cells)
photoreal/splat.ply	the 3DGS model moved into the simulation frame (for photoreal rendering in the sim)
frame/transform.json	sim_from_source transform and the evidence used to find the floor
manifest.json	bundle version, tier, stages and QA summary

---

🚀 Getting Started

Download one scene, or everything

pip install -U "huggingface_hub[cli]"

# One scene (about 22 GB)
hf download OverTheReality/OverMaps360_50 \
    --include "0313afa5-8b3c-4fb1-8245-c6b5abf738d2/*" \
    --repo-type dataset --local-dir ./OverMaps360-50

# Only the light assets of every scene (mesh, splat, IMU, poses, simulation bundle), no image archives
hf download OverTheReality/OverMaps360_50 \
    --exclude "*.tar" \
    --repo-type dataset --local-dir ./OverMaps360-50

Read the reconstruction

import pycolmap, numpy as np, pandas as pd

scene = "OverMaps360-50/0313afa5-8b3c-4fb1-8245-c6b5abf738d2"
rec = pycolmap.Reconstruction(f"{scene}/colmap_perspective/sparse/0")
print(rec.summary())                       # 11,500 images, ~1 M points

# camera centres in the metric scene frame (up = -Y)
centres = np.array([im.projection_center() for im in rec.images.values()])

# 1 kHz IMU and 1 Hz GPS
imu = pd.read_csv(f"{scene}/insta360_imu.csv")            # t_sec, gyro_*_dps, accel_*_mg
gps = pd.read_csv(f"{scene}/backgroundlocationData.txt", names=["epoch_s", "lat", "lon", "accuracy_m"])

Run the scene in MuJoCo

import mujoco
model = mujoco.MjModel.from_xml_path(f"{scene}/simulation/0313afa5.sre/scene.xml")
data = mujoco.MjData(model)
mujoco.mj_step(model, data)

---

📊 Manifest & Fields

dataset_manifest.parquet (and the identical dataset_manifest.csv) has one row per scene. Paths are relative to the dataset root.

Column	Description
mapping_id	Primary key, the scene UUID and folder name.
recording_360_path	Censored Insta360 recording (.insv).
imu_path	1 kHz IMU CSV.
location_path	1 Hz GPS log.
fisheye_lens0_path, fisheye_lens1_path	Fisheye frame archives (.tar).
perspective_images_path, perspective_masks_path, perspective_sparse_path	Perspective views archive, their masks archive and the COLMAP model folder.
masks_path	Archive of the privacy masks of fisheye frames and smartphone photos.
smartphone_raw_path, smartphone_poses_path, smartphone_audio_path	Smartphone photos archive, ARKit poses CSV, audio clips folder (empty when the scene has no audio).
training_cameras_path	Cameras of the 3DGS training.
gaussian_ply_path, gaussian_splat_path	3DGS model (PLY and .splat).
dense_ply_path	Dense point cloud.
mesh_glb_path, mesh_usdz_path	Textured mesh.
lod_high_path, lod_mid_path, lod_low_path, lod_json_path	Mesh LODs and their report.
simulation_path, certification_path, datasheet_path	Simulation bundle, certification reports, datasheet.
caption	VLM-generated natural-language description of the scene.
weather	Dominant weather label (Sunny, Overcast, Indoor, …).
time_of_day_algorithmic	Time-of-day bin inferred from lighting (Morning, Afternoon, …).
crowd_density	How many people were visible during the capture (Empty, Low, Moderate, High).
brightness	Lighting descriptor (Bright, Soft, Dim, …).
scene_type	Coarse region + fine-grained spot, semicolon separated (e.g. Urban;City Street).

Sensor file formats

File	Columns / format
insta360_imu.csv	t_sec (seconds from video start, can be negative for pre-roll), gyro_x/y/z_dps (°/s), accel_x/y/z_mg (milli-g), 1 kHz
backgroundlocationData.txt	epoch_s, lat, lon, accuracy_m, 1 Hz, no header
smartphone_poses_<seq>.csv	one row per smartphone photo: imageName, ARKit rotation quaternion (rot_w..rot_z) and position (pos_x..pos_z, metres, ARKit frame), timestamps, GPS fix, intrinsics (fx, fy, resW, resH, cx, cy), compass, satellites, speed, accelerometer and gyroscope at shot time
training_cameras.json	list of {id, img_name, width, height, position, rotation, fx, fy} (camera-to-world rotation as 3×3, 800×800 views)
gaussian_splatting.ply	standard 3DGS PLY: x y z nx ny nz f_dc_0..2 f_rest_0..44 opacity scale_0..2 rot_0..3
gaussian_splatting.splat	32 bytes per Gaussian: position (3×f32), scale (3×f32), colour RGBA (4×u8), rotation (4×u8)

---

✅ Certification

Each scene was certified before publication with ovr-maps-360-simkit, our open-source toolkit that builds the simulation bundle, the LODs and the datasheet from a scene and gates it. Every report in simulation/certification/ and datasheets/ can be reproduced with it (simkit regate simulation/<id8>.sre). The MuJoCo gate (simulation/certification/mujoco_physics_gate.json) runs 8 checks on the actual simulation bundle:

#	Check	Threshold
1	Up direction verified against the splat and the camera heights	≥ 90 % of cameras have ground below them
2	Scene is plumb (walls vertical)	tilt < 2°
3	Floor registration: navmesh vs walked floor, coverage of the walk	median < 5 cm, ≥ 20 % of cameras on walkable cells
4	Splat ↔ collision alignment residual	< 25 cm
5	Walkable area	≥ 5 m²
6	Probes falling through the collision geometry	0 %
7	Probes settled within 3 s	≥ 90 %
8	Contact penetration, p95	< 20 mm

The Isaac Sim report (isaac_drop_test.json) is a supporting drop test of a rigid sphere on a certified floor point, run headless with PhysX. All 50 scenes pass with tier T1a; measured values are in the JSON reports and in datasheets/<uuid>.json.

Across the 50 scenes: floor registration median 0.1 cm, plumb tilt median < 0.5°, walkable area median 3,548 m², 0 leaked probes.

---

⚖️ Licensing

This dataset is distributed under the Creative Commons Attribution-NonCommercial 4.0 International license.

• ✅ Research use: allowed and encouraged.
• ❌ Commercial use: prohibited.

Please attribute OverTheReality and link back to this repository when sharing derivatives.

⚠️ Disclaimer

This dataset is provided "as is" without warranties of any kind. While we strive for accuracy and quality, users assume all risks associated with its use. Over Holding Srl is not liable for any damages arising from its application. Despite rigorous privacy measures (segmentation-based masking of people and license plates applied to every published image and to the 360° video), we cannot guarantee the complete removal of all sensitive information. Users are responsible for ensuring compliance with privacy regulations when utilizing this dataset. If you identify any privacy concerns, please contact us immediately at data@ovr.ai.

@misc{OverMaps360_50,
  author = {OverTheReality},
  title = {{OverMaps360-50 Dataset}},
  howpublished = {Hugging Face Datasets},
  url = {https://huggingface.co/datasets/OverTheReality/OverMaps360_50},
  year = {2026},
}