01
Spherical imaging
A continuous surrounding observation captured at 5188 × 1979 pixels.
360° × 136.70°horizontal × vertical FoV with a DuxCam M4 spherical camera
Spherical 3D perception dataset · 2026
Spheriverse
3D Scene Understanding from
Spherical Observations in the Wild
89,674Spherical images
644Image–LiDAR sequences
13Diverse regions
30+Benchmarked methods
Dataset in motion
A continuous spherical drive.
Explore one continuous real-world sequence through its complete surrounding view—the native visual domain of Spheriverse and the starting point for its 3D perception tasks.
Dataset at a glance
Built for real-world spherical spatial perception.
Spheriverse combines high-quality spherical imagery with dense LiDAR observations and spatial annotations across geographically and visually diverse driving environments.
02
Dense LiDAR sensing
High-density range observations provide reliable geometric support for 3D annotation.
128 beamsHesai OT128 LiDAR
03
Synchronized acquisition
Camera, LiDAR, and GNSS/INS streams are aligned within a unified acquisition platform.
20 s · 5 Hzper synchronized sequence
Scene diversity
Five major scene categories.
Thirteen fine-grained environments.
Traffic function, environmental context, spatial layout, illumination, and weather vary substantially across the collection.
Urban Core and Mixed-Use Zones
1.1Central Functional ZoneSequence 294 · 5 s
1.2Mixed Residential–Commercial DistrictsSequence 526 · 5 s
Expressway Corridors and Peri-urban Areas
2.1Cross-Regional Expressway CorridorsSequence 44 · 5 s
2.2Urban–Rural Interface ZonesSequence 354 · 5 s
Functional and Restricted Operational Areas
3.1Waterfront AreasSequence 28 · 5 s
3.2Construction ZoneSequence 541 · 5 s
Structurally Constrained Transportation Areas
4.1Elevated Bridge EnvironmentsSequence 64 · 5 s
4.2Underground Confined EnvironmentsSequence 693 · 5 s
4.3Multi-Level Structural SpacesSequence 324 · 5 s
Rural and Natural Passage Environments
5.1Rural RoadsSequence 364 · 5 s
5.2Forest TracksSequence 437 · 5 s
5.3Village Access LanesSequence 380 · 5 s
5.4Farm Field TracksSequence 419 · 5 s
Semantic annotations
One dataset,
three levels of understanding.
The hierarchy preserves detailed real-world concepts while providing standardized category spaces for reproducible evaluation.
33Fine-grained attributes
→
14Intermediate classes
→
9Unified classes
Pedestrian
Vehicle
Cyclist
Building
Vegetation
Pole & Barrier
Road
Surface
Other
01 / TAXONOMY
From detailed attributes to unified evaluation classes
Percentages report the annotation share at each semantic level.
| Fine-grained attribute | Share | Intermediate class | Share | Unified class | Share |
|---|---|---|---|---|---|
| 01Pedestrian | 0.0520% | 1. Pedestrian | 0.0520% | 1. Pedestrian | 0.0520% |
| 02Passenger vehicle | 9.6620% | 2. Vehicle | 10.0105% | 2. Vehicle | 10.0105% |
| 03Heavy vehicle | 0.2452% | ||||
| 04Construction vehicle | 0.1033% | ||||
| 05Moving bicycle | 0.0015% | 3. Cyclist | 0.2242% | 3. Cyclist | 0.2242% |
| 06Stationary bicycle | 0.0066% | ||||
| 07Bicycle group | 0.0015% | ||||
| 08Moving motorcycle | 0.0559% | ||||
| 09Stationary motorcycle | 0.0828% | ||||
| 10Motorcycle group | 0.0028% | ||||
| 11Other cyclist | 0.0731% | ||||
| 12Permanent building | 15.7269% | 4. Building | 15.8154% | 4. Building | 18.0318% |
| 13Temporary building | 0.0885% | ||||
| 14Bridge | 0.2767% | 5. Bridge | 0.2767% | ||
| 15Tunnel | 1.2082% | 6. Tunnel | 1.2082% | ||
| 16Advertising object | 0.3067% | 12. Man-made | 0.5802% | ||
| 17Art object | 0.0298% | ||||
| 18Functional object | 0.2073% | ||||
| 19Other man-made object | 0.0364% | ||||
| 20Pavilion | 0.1526% | 13. Pavilion | 0.1526% | ||
| 21Vegetation | 26.6991% | 7. Vegetation | 26.6991% | 5. Vegetation | 26.6991% |
| 22Bollard | 2.4379% | 11. Barrier | 6.6195% | 6. Pole & Barrier | 7.3890% |
| 23Fence | 3.7391% | ||||
| 24Curb | 0.4326% | ||||
| 25Other barrier | 0.0099% | ||||
| 26Pole | 0.7695% | 8. Pole | 0.7695% | ||
| 27Drivable road | 28.4774% | 9. Road | 30.9190% | 7. Road | 30.9190% |
| 28Non-drivable road | 2.4416% | ||||
| 29Vegetated surface | 0.9616% | 10. Surface | 2.8002% | 8. Surface | 2.8002% |
| 30Farmland | 1.5752% | ||||
| 31Gravel | 0.1357% | ||||
| 32Other surface | 0.1277% | ||||
| 33Other | 3.8728% | 14. Other | 3.8728% | 9. Other | 3.8728% |
02 / SPATIAL COMPOSITION
Semantic composition changes with direction and distance
Each column is normalized to 100%; darker cells indicate a larger within-bin class share.
AzimuthRelative to the ego-vehicle heading
Radial distanceMeasured from the ego-vehicle origin
0%50%Within-bin class share
Unified benchmarks
Three tasks for spherical 3D scene understanding.
More than 30 representative methods are evaluated under common protocols with overall and scene-wise analyses.
Reference model
SphereOcc
An occupancy framework that integrates spherical geometry and semantic evidence into metric voxel representations through a geometry–semantics coupled cross-space mechanism.
13.91%mIoU+1.70 pp vs. prior best
24.65%GeoIoU+2.10 pp vs. prior best
Spherical geometry organization
Forms Cartesian–spherical voxel features by embedding sphere-aligned range–azimuth relations.
Semantic evidence re-querying
Uses RTZ-conditioned voxel queries to retrieve relevant evidence from source spherical image features.
Resources
Start exploring Spheriverse.
ManuscriptPaper PDF[arXiv]
ImplementationSource codeComing soon
Dataset releaseDownload dataComing soon
EvaluationBenchmark serverComing soon
Citation
Cite Spheriverse
@article{teng2026spheriverse,
title = {{Spheriverse:} {3D} Scene Understanding from Spherical Observations in the Wild},
author = {Teng, Fei and Wu, Sheng and Duan, Mengfei and Zhao, Guoqiang and Ma, Junhui and Luo, Kai and Li, Siyu and Shi, Hao and Li, Zhiyong and Yang, Kailun},
journal = {arXiv preprint arXiv:2609.09012},
year = {2026}
}