Datasets:

nvidia
/

video-to-data-object-assets

NVIDIA Video-to-Data Object Assets is a collection of textured 3D object meshes and simulation-oriented USD packages produced for the NVIDIA Video-to-Data project.

The repository is hosted as a Hugging Face dataset for versioned distribution, but its contents are 3D assets rather than raw recordings, an annotated machine-learning dataset, or a benchmark.

Version and contents

This is the V2D v0.3 object-asset release.

Content	Count
Distinct objects	69
Object/method packages	168
BundleSDF packages	43
Einstar packages	59
SAM3D packages	66

An object can have packages from more than one reconstruction method. Method coverage is not uniform, and packages from different methods should not be interpreted as equivalent ground truth.

Each package follows this layout:

<object_id>/
└── <method>/                 # bundlesdf, einstar, or sam3d
    ├── output.glb            # self-contained textured visual mesh
    ├── output.usd            # simulation-oriented package entry point
    ├── visual_asset.usd      # visual layer referenced by output.usd
    └── textures/             # present when referenced by the USD package
        └── *.png

Use output.usd as the USD entry point. Keep it together with visual_asset.usd and the adjacent textures/ directory. Use output.glb when a self-contained visual mesh is preferred.

Download

Install the Hugging Face CLI, then run:

hf download nvidia/video-to-data-object-assets \
  --type dataset \
  --local-dir ./video-to-data-object-assets

Individual objects or methods can be downloaded with --include, for example:

hf download nvidia/video-to-data-object-assets \
  --type dataset \
  --include "basketball/einstar/**" \
  --local-dir ./video-to-data-object-assets

Generation methods

BundleSDF

BundleSDF packages were reconstructed from calibrated object captures with the V2D BundleSDF workflow. The workflow estimates camera motion, obtains object masks and depth, reconstructs and aligns the observed surface, and exports a textured metric-scale mesh.

SAM3D

SAM3D packages were generated from selected segmented object views using the V2D SAM3D workflow, followed by scale and pose alignment. Because this method performs generative single-image 3D reconstruction, it can synthesize plausible surfaces that were not visible in the selected view.

Einstar

Einstar packages originated from in-house handheld-scanner captures. The exported scanner meshes were normalized and packaged through the same V2D simulation-asset workflow.

The reconstruction software, containers, model weights, raw captures, and intermediate outputs used by these workflows are not included in this dataset.

Scale and physical properties

The USD workflow targets meter-scale, Z-up rigid objects and authors collision geometry, mass, inertia, and a physics material. These values make the packages useful as simulation starting points; they are not measured physical ground truth.

Before using an asset in an experiment or application:

verify its dimensions and scale;
inspect visual geometry and collision geometry;
tune mass, inertia, friction, and restitution for the target scenario; and
confirm the intended support pose and dynamic behavior.

Validation

Every included object/method package contains the required GLB and USD entry-point files, along with referenced visual layers and textures where applicable. The release contains:

158 packages that passed the scoped simulation drop test;
7 packages that passed structural USD validation when no matching sequence was available for the behavioral test; and
3 spherical tennis-ball packages that passed simulation validation with the upright-standing check marked not applicable.

These checks are release filters, not comprehensive quality scores. Users should perform task-specific visual, geometric, collision, and dynamics validation.

Intended uses

The assets are intended for:

robotics simulation and manipulation research;
Video-to-Data reconstruction and grounding workflows;
object-pose and scene-reconstruction experiments;
synthetic observation generation; and
comparison of available reconstruction methods for an object.

The assets are not measured CAD, manufacturing references, safety-certified collision models, or a statistically representative machine-learning benchmark.

Limitations

Geometry, texture, scale, and material quality vary by object and method.
Reflective, transparent, dark, textureless, deformable, thin, or partially observed surfaces are difficult to reconstruct.
BundleSDF meshes can omit or thicken thin parts and can contain holes, fragments, or texture bleeding.
SAM3D meshes can contain synthesized unseen geometry and inaccurate proportions or scale.
Scanner meshes remain sensitive to scan coverage, occlusion, registration, and surface properties.
Collision geometry and physical properties are generated approximations.
Not every object is available from every reconstruction method.
Raw captures and complete per-run parameters are not included, so the assets cannot be independently regenerated from this repository alone.
Object identifiers are descriptive names and do not imply affiliation with or endorsement by any third party.