Datasets
Nova Datasets
- Street
- Open space
- Indoor
Scene structure
Instance-level decomposition.
Every part of a Nova scene is its own object, with a name, a semantic class, a material and collision geometry. Any part can be selected, moved or replaced.
Geometry and kinematics
3D bounding box and articulation annotations.
Each object carries a metric 3D bounding box, and each articulated part its joint type, pivot, axis and limits, all in one world frame. Every joint is verified in physics simulation to open to its limit.
Contributions
- 01
Instance-level decomposition
Every object is a separate, named part that can be selected, moved, replaced or removed.
- 02
Instance and semantic labels
Each object carries an instance ID and a semantic class, exact by construction.
- 03
Articulation annotations
Doors, gates and lids on revolute and prismatic joints, with axes, limits and damping.
- 04
Physical properties
Mass, inertia and friction on every moving body, each recorded with its source.
- 05
Collision geometry
Simulation colliders built by convex decomposition of each part.
- 06
Procedural scene programs
The Python program that builds the scene, for controlled variation and domain randomization.
- 07
Cross-simulator export
Native files for Isaac Sim, MuJoCo, Gazebo and PyBullet from one source.
- 08
PBR materials
Albedo, normal, roughness and ambient-occlusion maps for physically based rendering.
- 09
Capture-grounded inventory
Each reconstructed object is linked to the source images it was observed in.
- 10
Provenance and validation
Inputs, settings and versions for every scene, and every joint push-tested in simulation.
- Objects
- Articulated joints
- Semantic classes
| Objects | Articulated joints | Semantic classes |
|---|---|---|
| 43 | 0 | 9 |
| 50 | 0 | 9 |
| 55 | 1 | 10 |
| 66 | 1 | 10 |
| 67 | 1 | 10 |
| 73 | 1 | 11 |
| 79 | 2 | 13 |
| 79 | 2 | 15 |
| 135 | 2 | 15 |
| 141 | 3 | 16 |
| 144 | 3 | 17 |
| 160 | 4 | 18 |
| 166 | 4 | 20 |
| 167 | 4 | 21 |
| 172 | 4 | 21 |
| 202 | 4 | 23 |
| 205 | 4 | 23 |
| 221 | 5 | 24 |
| 231 | 5 | 25 |
| 231 | 5 | 26 |
| 232 | 5 | 26 |
| 247 | 6 | 26 |
| 256 | 6 | 27 |
| 320 | 7 | 28 |
| 325 | 7 | 29 |
| 334 | 7 | 31 |
| 343 | 8 | 31 |
| 352 | 10 | 33 |
| 369 | 10 | 34 |
| 386 | 10 | 35 |
| 424 | 10 | 35 |
| 508 | 13 | 36 |
| 572 | 14 | 36 |
| 586 | 15 | 36 |
| 586 | 17 | 37 |
| 603 | 18 | 38 |
| 622 | 19 | 38 |
| 629 | 21 | 42 |
| 718 | 27 | 48 |
| 896 | 31 | 58 |
| 1065 | 37 | 58 |
Collections
| Source | Setting | Size | Capture input | USD | MJCF | SDF | PyBullet |
|---|---|---|---|---|---|---|---|
| Scene datasets | |||||||
| WanderlandNYU AI4CE | Outdoor urban streets, New York and Jersey City | 20 scenes | Images (10) · Images + LiDAR mesh (10) | ✓ | ✓ | ✓ | ✓ |
| SceneSmith roomsSceneSmith benchmark | Indoor rooms | 20 scenes | Text brief | ✓ | ✓ | ✓ | ✓ |
| ScanNet++ indoorTU Munich | Indoor rooms | In progress | Images + LiDAR mesh | – | – | – | – |
| HarborsideNova capture | Office lobby, Jersey City | 1 scene | Images | ✓ | ✓ | ✓ | ✓ |
| Object datasets | |||||||
| Nova3D-BenchNova3D | Objects, six domains | 54 assets | Text or image prompt | – | – | – | – |