license: cc-by-4.0
task_categories:
- robotics
tags:
- video
- depth
- human
- object
- reconstruction
- locomotion
- manipulation
- demonstration
size_categories:
- 1K<n<10K
Dataset Description:
FORM-HOI (Foundry for Reconstruction from Multiview HOI) contains HOI video sequences paired with reconstructed human and object trajectories. Each sequence provides:
- RGB and depth videos from four calibrated views
- Mesh of object used for vision-based tracking
- Per-frame human pose, in SOMA and MHR formats
- Per-frame 6 DoF object pose
- Metric scale object mesh
This dataset is ready for commercial or non-commercial uses.
Dataset Owner(s):
NVIDIA Corporation
Dataset Creation Date:
July 2026
Version:
0.1.0
License/Terms of Use:
Creative Commons Attribution 4.0 International (CC BY-4.0)
Intended Usage:
This dataset is ideal for training HOI reconstruction models like CARI4D. It can also be used for grounding during robot policy training.
Dataset Characterization
** Data Collection Method
- Human - The HOI sequences are performed and recorded by human operators.
** Labeling Method
- Hybrid: human, automated - HOI sequences are ingested and processed by an automated pipeline to generate pose annotations. Human annotators label segments where pose annotations do not meet standards.
Dataset Format
| Property | Value |
|---|---|
| 4x RGB image streams | 1536x1152 RGB image arrays encoded with H.264 in MP4 container |
| 4x depth image streams | 768x576 depth image arrays compressed with HDF5 |
| 4x human mask image streams | 1536x1152 B/W image arrays compressed with HDF5 |
| 4x object mask image streams | 1536x1152 B/W image arrays compressed with HDF5 |
| Calibration | edex JSON-formatted values for per-camera intrinsics and extrinsics |
| Human poses (SOMA) | soma_params.npz compressed arrays of SOMA joint and shape parameters |
| Human poses (MHR) | mhr_params_mv.pt compressed arrays of MHR join and shape parameters |
| Object pose | poses.npy (N, 4, 4) array of per-fram object rigid-transforms |
| Object mesh | output_aligned.glb metric textured object mesh in GLB format |
| Object mesh symmetry | output_symmetry.json discrete/continuous rotational symmetries of the object mesh along principal axes |
| Ground plane | ground_plane.json 4 coefficients of the estimated ground plane, Ax+By+Cz+D=0 |
| Quality assessment results | failure_segments.json half-open ranges of frame indices flagged by human quality review or automated metric checks; see the explanation below |
failure_segments.json records intervals flagged by human quality review or automated metric checks:
- Human QC: Intervals where a reviewer visually inspected the rendered overlays and judged the pseudo-ground-truth annotations to be inaccurate.
- Chamfer distance: Intervals flagged by discrepancies between reconstructed geometry and observed depth. These discrepancies may reflect annotation errors or depth noise; in practice, depth noise is frequently the underlying cause.
- Silhouette bounding box containment: Intervals flagged by inconsistencies between projected geometry and segmentation-derived bounding boxes. These inconsistencies may reflect annotation errors or noisy segmentation masks.
Metric-based flags indicate potential quality issues and should not be interpreted as confirmed annotation errors.
Dataset Quantification
The statistics below cover only the cleaned release, excluding the held-out object sequences.
| Property | Value |
|---|---|
| Episodes | 4,135 |
| Total episode duration | ~29.144 hours |
| Unique object IDs | 61 |
| Unique person IDs | 22 |
| Downloadable archive size | ~4.504 TB |
Total duration counts each episode once, irrespective of the number of camera views. Archive size excludes release manifests and uses decimal TB.
Reference(s):
Technical report is in progress. Code is integrated in the video_to_data repo.
Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.