OverMaps · Data licensing
OverMaps: a real-world 3D dataset for AI labs
Posed multi-view imagery, 3D Gaussian Splatting, COLMAP geometry, depth and Qwen3-VL captions for world models and robot simulation. Test the free 1K sample on Hugging Face or license the full archive.
Captured via the Map2Earn contributor network · Need custom capture of your own facility instead? See custom data collection →
Real OverMaps scene: multi-view trajectory, COLMAP point cloud & 3D Gaussian Splatting
Every scene includes RGB frames, 4×4 metric SLAM poses, 3DGS (.ply), COLMAP sparse/dense reconstructions, depth, Qwen3-VL captions and YOLOv6 privacy masks.
Key dataset numbers (verified · OverMaps_172k & OverMaps-360Cam)
- 171,635
- Smartphone 3D scenes
- ~501 images/scene avg
- 86M+
- Multi-view JPEG images
- ~672 TB full archive
- 2,712
- 8K 360° walking scenes
- ~60 TB · 500+ hrs · 2,200 km
- 29 / 125
- Countries & cities
- Insta360 X5 8K captures
- 1,000
- Open sample on HF
- ~3.9 TB + 50 360Cam (~1.1 TB)
What’s inside
A 3D dataset for AI training: multi-view image dataset with camera poses and gaussian splatting dataset
Every scene in OverMaps is processed through a standardized spatial pipeline so AI labs can ingest aligned RGB frames, metric 6-DoF poses, neural rendering representations, geometry, and vision-language captions without building custom SfM or privacy-scrubbing infrastructure.
Modalities included per scene
| Criterion | Specification & pipeline | Training & research use |
|---|---|---|
| Multi-view RGB imagery + EXIF | High-resolution JPEG sequences (~501 images/scene on average in Smartphone; 8K 3840×3840 dual-fisheye in 360Cam) | Preserved camera intrinsics and EXIF metadata for multi-view geometry and world model pre-training |
| Metric camera poses (4×4) | Camera-to-world 4×4 transformation matrices at metric scale (ARKit / ARCore visual-inertial SLAM) | True metric scale without scale-ambiguity drift across indoor and outdoor trajectories |
| 3D Gaussian Splatting (3DGS) | Pre-computed 3D Gaussian Splatting radiance fields for every scene (.ply) | Photorealistic novel view synthesis, real-time rendering and sim-ready visual environments |
| COLMAP sparse & dense reconstructions | NetVLAD + ALIKED + SuperGlue + pixSfM pipeline (plus textured meshes and walkable navmeshes in 360Cam) | Structure-from-Motion baselines, geometric supervision, and navigable collision meshes for robot simulation |
| Depth maps | Per-frame aligned depth maps across captured trajectories | Monocular and multi-view depth estimation, 3D occupancy, and spatial grounding |
| Semantic annotations (Qwen3-VL) | Scene categories, weather, lighting conditions, crowd density, and dense per-image captions generated via Qwen3-VL | Vision-language-action (VLA), vision-and-language navigation (VLN), and text-to-3D / world model conditioning |
| Privacy-preserving pipeline (YOLOv6) | Automated detection and inpainting masks for people, faces, vehicles, and license plates via YOLOv6 | Privacy-scrubbed imagery and segmentation masks ready for commercial AI lab training pipelines |
Dataset variants & scale
| Criterion | Scenes, imagery & storage volume | Capture hardware & included modalities |
|---|---|---|
| OverMaps-Smartphone (OverMaps_172k) | 171,635 real-world scenes · 86M+ high-res JPEG images (~501/scene) · ~672 TB full archive | Consumer smartphones (ARKit/ARCore metric SLAM) via Map2Earn contributors; RGB, 4×4 poses, 3DGS, COLMAP, depth, Qwen3-VL captions, YOLOv6 privacy masks |
| OverMaps-360Cam | 2,712 walking scenes across 29 countries & 125 cities · 500+ hours · ~60 TB · 2,200 km of walked paths · 8K dual-fisheye (3840×3840) | Insta360 X5 walking captures; includes full 360° imagery, 3DGS, COLMAP, textured 3D meshes, and walkable navigation meshes (navmeshes); 50-scene open sample (~1.1 TB) |
| OverMaps_1k & 360Cam open samples | 1,000 Smartphone scenes (~3.9 TB) on Hugging Face + 50 360Cam scenes (~1.1 TB) | Free evaluation subsets so researchers can inspect every modality, camera pose format, and 3DGS/COLMAP output before licensing |
Note: The 171,635 scenes (86M+ images, ~672 TB) in OverMaps_172k and 2,712 scenes in OverMaps-360Cam represent the packaged OverMaps training dataset, distinct from the 270,000+ total locations mapped on the OVER visual positioning platform.
Try before you license
Inspect OverMaps_1k on Hugging Face, real scene captures and the Technical Report
We designed the evaluation workflow for ML researchers: download the 1,000-scene curated subset (~3.9 TB) on Hugging Face, verify the file layout and pose accuracy in your own training loader, and review the full technical report before involving procurement.
Indoor multi-room capture
Multi-view trajectory with metric 4×4 camera poses, sparse/dense point cloud and pre-computed 3D Gaussian Splatting.
Outdoor architectural scene
Real-world outdoor capture with preserved EXIF, YOLOv6 privacy masks, aligned depth maps and Qwen3-VL scene captions.
Urban street-level environment
Dense street-level loop (~501 images/scene average) for visual localization, world models and novel view synthesis.
8K 360° walking trajectory + navmesh
Full-sphere 8K (3840×3840) walking capture with textured mesh and walkable navmesh for embodied AI and robot simulation.
Verification resources for researchers
Direct links to data, benchmarks & interactive samples
- Hugging Face (
OverTheReality/OverMaps_1k): huggingface.co/datasets/OverTheReality/OverMaps_1k ↗ — 1,000 curated scenes (~3.9 TB) with all 7 modalities. - Dataset Technical Report (PDF): Dataset_technical_report.pdf ↗ — sensor specs, NetVLAD + ALIKED + SuperGlue + pixSfM pipeline, 3DGS and Qwen3-VL annotation details.
- OVR-MAPS project page: ovr-platform.github.io/OVR-MAPS ↗ — schema documentation and modality overview.
- Interactive 3D scene explorer: marketplace.ovr.ai ↗ — inspect live 3D Gaussian Splatting scenes in the browser.
BibTeX citation
Cite OverMaps in your paper or technical report
@misc{OverMaps1k,
author = {OverTheReality},
title = {{OverMaps-1K Dataset}},
howpublished = {Hugging Face Datasets},
url = {https://huggingface.co/datasets/OverTheReality/OverMaps_1k},
year = {2025}
}Dataset created and maintained by Over Holding Srl (OVER). Full commercial and research training licenses for OverMaps_172k and OverMaps-360Cam are scoped on request.
Who uses it
World model training data and robot simulation dataset for AI labs
OverMaps is built for AI research labs, foundation model teams, and embodied AI groups that need real-world 3D diversity across indoor and outdoor scenes rather than synthetic renders or street-only driving logs.
01 · World models & spatial foundation models
World model training data at 172K-scene scale
Train generative world models, video-to-3D networks, and spatial foundation models on 86M+ real-world multi-view images paired with metric 4×4 camera trajectories, depth maps, and dense Qwen3-VL captions across diverse lighting and weather conditions.
02 · Embodied AI & robot simulation
Robot simulation dataset with 3DGS & walkable navmeshes
Import pre-computed 3D Gaussian Splats, textured meshes, and walkable navmeshes from OverMaps-360Cam (2,712 scenes and 2,200 km of walked paths across 29 countries and 125 cities) into simulation pipelines for VLN, VLA, and locomotion policy pre-training.
03 · Novel view synthesis & 3D reconstruction
Gaussian splatting dataset & Structure-from-Motion ground truth
Benchmark feed-forward 3DGS, NeRF, and dense reconstruction models against 171,635 scenes processed with NetVLAD, ALIKED, SuperGlue, pixSfM, and COLMAP sparse and dense reconstructions.
04 · Visual Positioning Systems (VPS) & 6-DoF localization
Multi-view image dataset with camera poses at metric scale
Train and evaluate learned local descriptors, place recognition (NetVLAD), and 6-DoF visual relocalization under real-world viewpoint, illumination, and crowd-density variations.
For your procurement team
Evaluating OverMaps as a researcher? Share this summary with your licensing buyer
In most AI labs, an ML researcher or research engineer evaluates OverMaps_1k on Hugging Face first, then brings in procurement or legal to license OverMaps_172k or OverMaps-360Cam. Here are the verified facts your buyer needs:
- Rights holder & licensor
- Over Holding Srl (OVER), Viale Tricesimo 200, 33100 Udine, Italy.
- What can be licensed
- Full
OverMaps_172k(~672 TB, 171,635 scenes),OverMaps-360Cam(2,712 8K scenes, ~60 TB, 2,200 km walked), or custom subsets by modality, geography, or volume — plus ongoing Map2Earn drops. - Provenance & privacy pipeline
- Captured via the OVER Map2Earn contributor network on consumer mobile devices and 360° cameras; processed with YOLOv6 face, person, vehicle, and license plate detection and inpainting masks.
- License scoping & delivery
- Terms are scoped on request based on volume and use case. Submit the form below; we reply by email to scope the license.
Two ways to work with OVER Robotics & Data
Dataset licensing vs. custom site capture
- 1. OverMaps data licensing (this page):
For AI labs and research teams training foundation models, world models, and robot simulation on our pre-collected global dataset (171,635 Smartphone scenes + 2,712 8K 360° scenes).
- 2. Custom data collection & VLN/VLA training:
For corporate and industrial teams that need on-demand 3D capture of their own warehouses, plants, or hospitals, sim-ready digital twins, and custom VLN/VLA policy training.
Go to Custom data collection & model training →
What is in the OverMaps dataset?
OverMaps by Over Holding Srl provides 171,635 real-world scenes and over 86 million high-resolution images (~672 TB in OverMaps_172k) plus 2,712 8K dual-fisheye 360° walking captures across 29 countries and 125 cities. Every scene ships with multi-view JPEG frames, metric 4×4 camera poses, 3D Gaussian Splatting, COLMAP reconstructions, depth maps, Qwen3-VL captions, and YOLOv6 privacy masks.
How do AI labs license OverMaps?
AI labs and corporate research teams license either the complete OverMaps_172k archive (~672 TB), the OverMaps-360Cam 8K walking collection, or a custom subset filtered by geography, modality, or scene category. Researchers typically evaluate the 1,000-scene sample first, then submit the form below with their procurement team; we reply by email to scope the license terms.
Can we try the dataset before buying?
Yes. Research teams can download OverMaps_1k directly from Hugging Face Datasets at no cost to inspect 1,000 curated real-world scenes (~3.9 TB) with full imagery, metric camera poses, 3D Gaussian Splats, COLMAP reconstructions, depth maps, and Qwen3-VL semantic annotations. You can also read the Technical Report PDF and explore interactive scene samples before licensing.
Questions buyers ask
How is the OverMaps dataset different from the 270,000+ VPS locations on the OVER platform?
The 270,000+ figure refers to all locations mapped across the OVER visual positioning platform. OverMaps is the curated, multi-modal AI training dataset packaged for licensing: 171,635 Smartphone scenes (86M+ images, ~672 TB in OverMaps_172k) and 2,712 8K 360Cam walking scenes (500+ hours, ~60 TB, 2,200 km of walked paths across 29 countries and 125 cities), with all 7 modalities pre-computed.
Can we license a subset by modality, country, or scene category instead of all 672 TB?
Yes. You can license the full OverMaps_172k archive (~672 TB), the OverMaps-360Cam 8K collection, or a filtered subset selected by scene category, lighting/weather tags (from Qwen3-VL annotations), geography, or specific modalities (for example RGB + poses + depth only, or 3DGS + navmeshes).
How is privacy handled in the imagery?
Every scene runs through an automated privacy-preserving pipeline based on YOLOv6 that detects people, faces, vehicles, and license plates, providing both inpainted imagery and segmentation masks.
Do you also capture custom sites on demand for a specific robot deployment?
Yes. OverMaps licensing covers our pre-collected global dataset for AI labs. If you need on-demand 3D capture of your own warehouses, factories, or hospitals—or custom VLN/VLA policy training on your facility—see our Custom data collection & model training service on /robotics-data.
Request a dataset license
Whether you are a researcher evaluating OverMaps_1k or a procurement lead scoping a full or subset license, tell us what you train and the volume you need. We reply by email to scope the license.