docs: finalize v0.3 object asset card

Document asset contents, download and usage, generation methods, validation, limitations, license, attribution, and support. No asset payload or visibility changes.

README.md

@@ -1,295 +1,176 @@
 ---
 pretty_name: NVIDIA Video-to-Data Object Assets
+license: cc-by-4.0
 task_categories:
-- robotics
+  - robotics
 size_categories:
-- n<1K
+  - n<1K
 tags:
-- 3d
-- robotics
-- object-reconstruction
-- simulation
-- glb
-- usd
-- video-to-data
+  - 3d
+  - robotics
+  - object-reconstruction
+  - simulation
+  - glb
+  - usd
+  - video-to-data
 ---
 
 # NVIDIA Video-to-Data Object Assets
 
-> **Draft release card:** This repository is private and under release review.
-> The asset license, public-release approval, and final Video-to-Data release
-> tag are not yet assigned. Do not make the repository public until the items
-> in [Release requirements](#release-requirements) are complete.
-
-## Current legal-review status
-
-An object-level visual review was completed for all 69 objects using raw object
-photos and available asset evidence. Recognizable logos or trademarks were
-identified for 37 objects; no recognizable logo or trademark was identified
-for 32 objects. No objects remain unresolved at the object-identification
-level.
-
-Release-asset remediation was completed for the reviewed v0.3 scope.
-Recognizable marks in 48 EINSTAR or BundleSDF object/method packages were
-removed from `output.glb` and the corresponding external texture file (96
-payload files total). The related `output.usd` and `visual_asset.usd` files
-were unchanged because the remediation modified only visual content. Reviewed
-SAM3D packages had no readable logo or trademark requiring remediation.
-
-These technical changes do not constitute NVIDIA Product Legal approval. The
-repository must remain private until Product Legal approves the final asset
-set and asset license. The item-level source-object and brand mapping is
-maintained in NVIDIA internal release-review documentation and is
-intentionally not reproduced in this card, which is expected to become public.
-
-## Summary
-
-NVIDIA Video-to-Data Object Assets is a curated collection of textured 3D
-object meshes and simulation-oriented USD packages. The collection contains
-derived 3D assets, not raw recordings, an annotated machine-learning dataset,
-or a train/evaluation benchmark. It is hosted as a Hugging Face dataset
-repository to provide versioned distribution of the asset packages.
-
-The assets were produced from physical-object captures used by the NVIDIA
-Video-to-Data (V2D) project. Depending on the package, geometry came from the
-V2D BundleSDF workflow, the V2D SAM3D workflow, or an Einstar scanner capture.
-The resulting meshes were normalized and packaged for downstream use by the
-V2D HOI object-reconstruction and mesh-to-USD workflows.
-
-| Property | Release-candidate value |
+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](https://github.com/nvidia-isaac/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 |
-| Payload files | 606 |
-| Payload size | 4,050,613,707 bytes (approximately 4.05 GB) |
 | BundleSDF packages | 43 |
 | Einstar packages | 59 |
 | SAM3D packages | 66 |
-| Initial Hub staging date | 2026-08-13 |
 
-One object can have packages from more than one method. Method coverage is not
-uniform, and the alternatives should not be interpreted as equivalent ground
+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.
 
-## Project and version information
+Each package follows this layout:
 
-- **Project:** [NVIDIA Video-to-Data](https://github.com/nvidia-isaac/video_to_data)
-- **Reconstruction module:** [v2d_hoi_object_reconstruction](https://github.com/nvidia-isaac/video_to_data/tree/main/reconstruction/modules/v2d_hoi_object_reconstruction)
-- **Documentation:** [Video-to-Data reconstruction documentation](https://nvidia-isaac.github.io/video_to_data/reconstruction/)
-- **Dataset version:** V2D v0.3 release candidate
-- **Exact public source tag or commit:** **TBD before public release**
-- **Creator and curator:** NVIDIA Video-to-Data team
-- **License:** **TBD after NVIDIA Product Legal review**
+```text
+<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
+```
 
-The software licenses in the Video-to-Data source repository do not
-automatically license this asset collection. Before public release, this card
-and repository must include the approved asset license and any required
-third-party notices.
+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.
 
-## Generation methods
+## Download
 
-### BundleSDF
+Install the [Hugging Face CLI](https://huggingface.co/docs/huggingface_hub/guides/cli), then run:
 
-BundleSDF packages are reconstructed from calibrated stereo object scans. The
-V2D workflow estimates camera motion, detects and segments the prompted object,
-estimates stereo depth, reconstructs a textured implicit surface, aligns scan
-stages, and exports a metric-scale mesh. The workflow uses components including
-CuSFM, Grounding DINO, SAM2, FoundationStereo, BundleSDF, and FoundationPose.
+```bash
+hf download nvidia/video-to-data-object-assets \
+  --type dataset \
+  --local-dir ./video-to-data-object-assets
+```
 
-BundleSDF reconstructs observed geometry. Thin or incompletely observed parts
-can be missing or thickened, and outputs can contain holes, fragments, or local
-texture bleeding.
+Individual objects or methods can be downloaded with `--include`, for example:
 
-### SAM3D
+```bash
+hf download nvidia/video-to-data-object-assets \
+  --type dataset \
+  --include "basketball/einstar/**" \
+  --local-dir ./video-to-data-object-assets
+```
 
-SAM3D packages are generated from representative segmented views selected from
-a calibrated object capture. V2D runs single-image 3D reconstruction for the
-selected views, estimates scale and pose against silhouettes and optionally
-depth, ranks the candidates, and exports the selected result.
+## Generation methods
 
-SAM3D can generate plausible surfaces that were not visible in the source
-view. Visual completeness is therefore not evidence that unseen geometry,
-dimensions, orientation, or metric scale are accurate.
+### BundleSDF
 
-### Einstar
+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.
 
-Einstar packages originate from hardware-scanner captures rather than the
-BundleSDF or SAM3D mesh-generation stages. Scanner meshes were selected as
-reference assets and then normalized and packaged through the V2D
-simulation-asset workflow. Scanner hardware, software, and capture settings are
-not included in this release candidate and should be documented if exact
-scanner reproducibility is required.
+### SAM3D
 
-## Source data and provenance
+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.
 
-Before Hugging Face staging, the derived asset packages were stored in NVIDIA
-internal object storage under a release-controlled V2D mesh catalog. The Hub
-upload preserves each selected path relative to that catalog.
+### Einstar
 
-This release contains only selected derived files. It does **not** contain:
+Einstar packages originated from in-house handheld-scanner captures. The exported scanner meshes
+were normalized and packaged through the same V2D simulation-asset workflow.
 
-- raw stereo images, videos, or scanner projects;
-- camera calibration, depth maps, masks, bounding boxes, or object prompts;
-- human pose, hand pose, or object trajectory recordings;
-- reconstruction model weights;
-- intermediate reconstruction outputs, debug renders, or private validation
-  evidence; or
-- per-run generation reports and exact reconstruction parameters.
+The reconstruction software, containers, model weights, raw captures, and intermediate outputs
+used by these workflows are not included in this dataset.
 
-Because raw inputs and complete per-run parameters are not included, the
-released assets cannot be reproduced from this repository alone. The final
-V2D source release tag should be recorded above to identify the corresponding
-workflow implementation.
+## Scale and physical properties
 
-## Repository structure
+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.
 
-Each object/method package follows this layout:
+Before using an asset in an experiment or application:
 
-```text
-<object_id>/
-└── <method>/                 # bundlesdf, einstar, or sam3d
-    ├── output.glb            # canonical 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
-```
+- 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.
 
-Use `output.usd` as the USD package entry point. It is not a standalone file:
-keep `output.usd`, `visual_asset.usd`, and any adjacent `textures/` directory
-together. `output.glb` is the self-contained visual-mesh alternative.
+## Validation
 
-## Scale and physical properties
+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.
 
-The USD preparation workflow targets meter-scale, Z-up rigid objects and
-authors collision geometry, mass, inertia, and a physics material. These
-properties are intended to make the packages usable as simulation starting
-points; they are not a claim of measured physical truth.
-
-- Verify geometry dimensions and scale before use.
-- Treat authored mass, inertia, friction, and restitution as simulation
-  assumptions unless independently measured for the physical object.
-- The V2D converter uses a 0.3 kg mass and generic material values by default
-  unless explicitly overridden during package generation.
-- Collision geometry is an approximation and may not preserve small features,
-  cavities, thin parts, or contact surfaces.
-- Passing a simulation drop test does not establish manufacturing, metrology,
-  safety, or real-world dynamics accuracy.
-
-## Curation and validation
-
-An object/method package was eligible for staging only when the release
-manifest contained all three required files: `output.glb`, `output.usd`, and
-`visual_asset.usd`. Referenced texture files were included with the package.
-The staged payload was verified file-for-file and size-for-size against the
-release manifest before this card was drafted.
-
-The 168 included packages have the following release-candidate validation
-statuses:
-
-| Status | Packages | Meaning |
-|---|---:|---|
-| `drop_test_passed` | 158 | The prepared package passed the scoped simulation drop-test decision used for this release. |
-| `structural_validation_passed_sequence_unavailable` | 7 | USD structure passed validation, but a matching sequence was unavailable for the behavioral drop test. This is not a behavioral pass. |
-| `simulation_passed_upright_check_not_applicable` | 3 | Simulation validation passed, while the upright-standing check was not applicable to the spherical tennis-ball asset. |
-
-Some candidate packages that did not meet the scoped geometry or simulation
-validation criteria were excluded. Method coverage is therefore uneven, and
-an absent object/method package should not be interpreted as an upload error.
-
-Validation decisions are release filters, not comprehensive quality scores.
-Users should visually inspect the selected method and validate scale,
-collision, support pose, and dynamics for their own task.
+These checks are release filters, not comprehensive quality scores. Users should perform
+task-specific visual, geometric, collision, and dynamics validation.
 
 ## Intended uses
 
-Subject to the final license, these assets are intended for:
+The assets are intended for:
 
 - robotics simulation and manipulation research;
-- testing or demonstrating Video-to-Data reconstruction and grounding
-  workflows;
-- object-pose, scene-reconstruction, and simulation-pipeline experiments;
-- comparing reconstruction methods on the same broad object catalog; and
-- generating synthetic observations after task-specific asset validation.
+- 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 collection is not intended to serve as measured CAD, a dimensional
-reference, a manufacturing model, a safety-certified collision model, or a
-benchmark with statistically representative train/test splits.
+The assets are not measured CAD, manufacturing references, safety-certified collision models, or
+a statistically representative machine-learning benchmark.
 
 ## Limitations
 
-- Geometry, texture, physical scale, and material quality vary across objects
-  and reconstruction methods.
-- Reflective, transparent, dark, textureless, deformable, thin, or partially
-  observed surfaces are particularly difficult to reconstruct.
-- BundleSDF can omit or thicken thin parts and can introduce holes, fragments,
-  or texture bleeding.
-- SAM3D can synthesize unseen geometry and may produce inaccurate proportions
-  or metric scale.
-- Scanner outputs remain sensitive to scan coverage, occlusion, registration,
-  and surface properties.
-- Physics properties and colliders are generated approximations, not measured
-  ground truth.
-- Method availability is uneven: not every object has all three methods.
-- No raw capture or complete generation provenance is included for independent
-  reconstruction auditing.
-- Object names are descriptive identifiers and do not constitute a controlled
-  taxonomy.
-
-## Privacy, trademarks, and responsible release
-
-The payload is intended to contain only reconstructed object geometry,
-materials, and simulation packaging. It does not include raw recordings of
-people, speech, or human-motion annotations. Nevertheless, textured
-reconstructions can reproduce text, logos, trade dress, or other appearance
-from the captured physical objects.
-
-Object and product names are used only for identification. Their inclusion
-does not imply affiliation with or endorsement by any third-party brand owner.
-Before public release, reviewers should confirm that textures contain no
-incidental personal information and that the approved license and notices
-cover the captured objects, reconstructed geometry, textures, scanner outputs,
-and model-generated outputs.
-
-## License
-
-**No asset license has been assigned in this draft.** Public access must remain
-disabled until NVIDIA Product Legal approves the release and a license is added
-both to this card's metadata and to the repository. Access to the private
-release candidate does not itself grant redistribution or other public-use
-rights.
-
-The reconstruction and mesh-to-USD tools are used only to generate and validate
-the outputs and are not included in this repository. Model weights and prebuilt
-containers are also excluded. Their use is reviewed through NVIDIA's internal
-software and model compliance process; the final asset license applies only to
-the released output files.
-
-## Release requirements
-
-Complete the following before making the repository public:
-
-- [ ] Obtain NVIDIA Product Legal approval for the asset collection.
-- [ ] Confirm rights and required notices for source captures, branded objects,
-      scanner outputs, and generated textures and geometry.
-- [ ] Obtain NVIDIA Product Legal confirmation that the remediated payload is
-      acceptable, or exclude any assets that are not approved for release.
-- [x] Complete technical logo/trademark review for the reviewed v0.3 release
-      scope, including SAM3D assets.
-- [x] Replace the 48 affected EINSTAR/BundleSDF GLB and external texture pairs
-      with reviewed remediated copies and verify the resulting Hub content.
-- [ ] Confirm the release treatment of BundleSDF, SAM3D, and their model outputs.
-- [ ] Select the asset license, add a `LICENSE` file, and add the corresponding
-      `license` field to this card's YAML metadata.
-- [ ] Record the final V2D v0.3 public tag or commit under Project and version
-      information.
-- [ ] Confirm the public maintainer/contact route.
-- [ ] Review textures for incidental personal information.
-- [ ] Re-run payload and dependency verification on the final Hub revision.
-- [ ] Replace this draft notice and mark the completed items above.
+- 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.
+
+## License and attribution
+
+This dataset is licensed under the
+[Creative Commons Attribution 4.0 International License](https://creativecommons.org/licenses/by/4.0/)
+(CC BY 4.0).
+
+When using or redistributing these assets, attribute:
+
+> NVIDIA Corporation, *NVIDIA Video-to-Data Object Assets*, v0.3,
+> https://huggingface.co/datasets/nvidia/video-to-data-object-assets
+
+## Related resources
+
+- [Video-to-Data source repository](https://github.com/nvidia-isaac/video_to_data)
+- [HOI object-reconstruction module](https://github.com/nvidia-isaac/video_to_data/tree/main/reconstruction/modules/v2d_hoi_object_reconstruction)
+- [Reconstruction documentation](https://nvidia-isaac.github.io/video_to_data/reconstruction/)
 
 ## Maintenance and feedback
 
-This collection is maintained by the NVIDIA Video-to-Data team. Until a
-dataset-specific contact is approved, report technical issues through the
+The NVIDIA Video-to-Data team maintains this collection. Report technical issues through the
 [Video-to-Data GitHub issue tracker](https://github.com/nvidia-isaac/video_to_data/issues).