README.md

---
license: cc-by-nc-4.0
language:
- en
pretty_name: HOI4D VITRA Streaming v1
size_categories:
- 1K<n<10K
tags:
- robotics
- vision-language-action
- egocentric-video
- mano
- hand-object-interaction
---

# HOI4D → VITRA streaming release

This directory builds a strict, camera-relative VLA training release from the
official HOI4D RGB-D, camera, motion-mask, action, and MANO annotations.  The
target Hugging Face repository is
[`LeoJiangOR/hoi4d-vitra-streaming-v1`](https://huggingface.co/datasets/LeoJiangOR/hoi4d-vitra-streaming-v1).

## What makes the conversion difficult

| Issue | What is present in HOI4D | Training-safe treatment |
|---|---|---|
| Moving egocentric camera | MANO pose and translation are expressed in each frame's instantaneous camera, while `3Dseg/output.log` contains 300 Open3D camera extrinsics. | Parse the matrices as absolute `world → camera` poses, move every valid hand pose into one shared world frame, and let VITRA transform a future chunk into its current anchor camera. Never mix future camera frames. |
| MANO `trans` is not the wrist | Official forward is `ManoLayer(poseCoeff, beta) / 1000 + trans`. On checked examples, canonical joint 0 is about 86–105 mm from the model origin in `trans`. | Forward MANO and store canonical joint 0 as VITRA `transl`; retain source `trans` only as provenance. Both 51-D angle and 69-D keypoint modes then have the same wrist convention. |
| Separately licensed model | HOI4D publishes MANO parameters but not redistributable MANO model files. | Require local MANO v1.2 `MANO_LEFT.pkl` and `MANO_RIGHT.pkl`, record their hashes/configuration, and never upload the model files. Reproject forwarded joints and require agreement with source `kps2D`. |
| Partial hand release | The hand ZIP contains 1,858 sequence keys, but only 1,684 intersect the 2,973 RGB sequences. The strict official `release.txt` intersection is 1,682 sequences. Hand labels cover only nine categories (`C1,C2,C3,C5,C7,C12,C13,C14,C20`), and frame IDs are variable with real boundary and interior holes. | Join by the complete seven-token key and exact integer frame ID. Preserve missing frames as invalid loss masks; never use array-offset, nearest-frame, or interpolation to fabricate source GT. RGB-only and hand-only sequences receive quarantine records. |
| 15 Hz source vs 10 Hz controller | RGB, depth, camera, and hand labels are native 15 Hz; the target VLA control rate is 10 Hz with a 16-step action head. | Keep the source timeline unchanged. At training time, derive targets at `anchor + 0.1 … 1.6 s`: translation/keypoints use linear interpolation and rotations use SLERP. Do not persist 16 fake frames and do not add random noise to GT. Per-step masks require valid bracketing source labels. |
| Alignment cannot be inferred from filenames alone | A correct path join does not prove that a MANO fit lies on the corresponding RGB/depth frame. | Use three gates: (1) MANO-forward projection vs source-rendered `kps2D`; (2) projected hand vs independently annotated RGB motion mask; and (3) projected 3-D depth vs aligned 16-bit depth. Generate overlays and retain all raw metrics. |
| Action JSON is inconsistent with RGB duration | Two action schemas occur, and many files report 10 s while RGB is 20 s. | Normalize both schemas as a fine-action sidecar. Use the official `(category, task)` collection definition as the high-level instruction over the episode; do not turn uncertain action timestamps into VLA instruction windows. |
| Object annotation failure modes | Public HOI4D reports include filename/internal-frame mismatches and CAD/part alignment problems. | V1 is hand-action-first. Object JSON count/provenance is audited, but object transforms are not a training gate or silently promoted to trusted VLA labels. |
| Split leakage | The pinned release couples each participant with one camera (`H1↔camera01` through `H4↔camera04`). Random episode splitting leaks both person and camera appearance. | Group split: `H1+H2=train`, `H3=val`, `H4=test`. This is intentionally harder and keeps participant/camera identity disjoint. |

## Pinned source inventory

All large files are accepted only after matching official OneDrive object size
and SHA-256.  Exact mirrors are transport fallbacks, not alternative data
sources.  See [`source_manifest.json`](source_manifest.json).

The pinned RGB ZIP has 2,973 exact sequences: 1,022 from H1, 695 from H2,
532 from H3, and 724 from H4.  The official repository's `release.txt` lists
2,971 of them; two archive-only paths are preserved in audit metadata but are
quarantined by default.  Exactly 2,972 RGB sequences have 300 motion masks and
one RGB sequence has none.  RGB, action JSON, camera trajectory, and object
pose directories otherwise join one-to-one by the complete sequence key.

The hand ZIP has 458,739 right-hand and 31,845 left-hand pickle files over
1,858 keys.  Left-hand annotations occur only for chair (`C20`) sequences.
After exact intersection with official released RGB, the pre-geometry ceiling
is 1,682 sequences:

| Split | Participants/cameras | Pre-geometry episodes |
|---|---|---:|
| train | H1/camera01 + H2/camera02 | 976 |
| val | H3/camera03 | 267 |
| test | H4/camera04 | 439 |

These are inventory ceilings, not claimed final accepted counts.  The dataset
card is finalized only after full RGB-mask, depth, temporal, loader, manifest,
and immutable-remote verification.

## Output episode schema

Each accepted episode contains the original 300-frame H.264 MP4 plus one
VITRA episodic `.npy`.  The label stores:

- exact source frame IDs and 15 Hz timestamps;
- camera intrinsics and 300 `world → camera` extrinsics;
- per-side MANO beta, global/world and global/camera rotations, 15 local joint
  rotations, wrist translations, 21 joints in world/camera coordinates;
- source MANO parameters/model-origin translations and rendered `kps2D` for
  provenance;
- separate source-availability, reprojection, RGB-mask, geometric-visibility,
  depth, and training masks;
- exact-time 10 Hz future-step validity for a 16-step action head;
- official task instruction and normalized fine-action sidecar;
- source member identities, hashes/CRC, split group, coordinate semantics, and
  license metadata.

`joints_worldspace` is the canonical representation.  For an anchor frame
`a`, every future world joint `J_world(t)` must be transformed with the same
anchor extrinsic `E_a`:

```text
J_anchor(t) = E_a · J_world(t)
```

Applying `E_t` separately to every future frame would recreate the moving
camera and is not a valid action chunk.

## Training with VITRA

The accompanying loader patch is exported as
`pipeline/vitra/timestamp_resampling.py`; the source repository also routes
`hoi4d_vitra_train`, `hoi4d_vitra_val`, and `hoi4d_vitra_test` through
`FrameDataset`.  A standard 16-action configuration is:

```python
dataset = FrameDataset(
    dataset_folder="/path/to/hoi4d-vitra-streaming-v1",
    dataset_name="hoi4d_vitra_train",
    action_past_window_size=0,
    action_future_window_size=15,
    action_type="angle",       # 51 dimensions per hand
    # action_type="keypoints", # 69 dimensions per hand
    use_rel=False,
    rel_mode="step",
)
```

The first predicted row is at `observation_time + 0.1 s`, and the final row is
at `observation_time + 1.6 s`.  The loader uses the observation frame's camera
for the whole chunk, linearly interpolates Euclidean quantities, and performs
pairwise shortest-path SLERP for global and finger rotations.  It never treats
the next native 15 Hz frame (`+66.7 ms`) as one 10 Hz control step.

All three split-specific statistics filenames are provided for plug-and-play
loading, but they deliberately contain moments computed from the **train split
only**.  This avoids validation/test leakage.  Both angle and keypoint modes
are exercised at the first, middle, and last valid anchor of every accepted
episode before upload.

## Quality and scope

An episode enters an index only after all of these gates pass:

1. exact seven-token RGB/annotation/hand join and official-release membership;
2. exact 300-frame RGB/camera timeline and byte-preserved source video;
3. MANO-forward reproduction of source `kps2D`;
4. geometric visibility and independent RGB motion-mask alignment;
5. aligned 16-bit depth identity/alignment audit;
6. temporal wrist/joint/angular-velocity checks;
7. at least 64 final right-hand frames and 10 valid 10 Hz anchors; and
8. real VITRA loader smoke tests in both 51-D and 69-D modes.

Malformed individual hand pickle records are retained in provenance as
`source_member_present=True, source_invalid_record=True` but are masked from
loss at their original frame IDs.  Frames are never shifted, nearest-filled,
or silently synthesized.  Rejected episodes upload their audit report and
overlay but not training video/labels; source-only exclusions appear in the
global quarantine ledger.

This is a hand-action release, not a complete object-state benchmark.  HOI4D's
high-level collection tasks are useful commands but are coarser than narrated
robot demonstrations; fine action JSON and object poses remain sidecars rather
than trusted supervision.  No tactile signal is present.

## Reproducible streaming pipeline

- `tools/download_verified_sources.py`: resumable concurrent downloads with
  official byte-size and SHA-256 gates.
- `tools/download_onedrive_share.js`: keeps the anonymous OneDrive capability
  token inside Playwright and validates the downloaded file; tokens are never
  printed or persisted.
- `tools/inventory_hoi4d_archives.py`: safe ZIP inventory and complete-key join.
- `preprocessing/convert_hoi4d_episode.py`: per-episode geometry, video,
  language, RGB-mask, and MANO conversion.
- `geometry.py`: Open3D pose parsing and exact-time 15→10 Hz interpolation.
- `training_artifacts.py`: exact loader parity checks and mergeable float64
  normalization moments.
- `streaming/stream_hoi4d_to_hf.py`: SQLite claims, multi-worker conversion,
  one-pass split-depth audit, final gates, verified batched upload, and indices.
- `streaming/verify_hoi4d_release.py`: immutable remote tree, size/hash,
  manifest, ledger, and index verification.
- `tests/`: synthetic invariants plus a real-data pilot gate.

The source archives are downloaded in parallel while storage permits.  Episode
outputs are uploaded in bounded batches, checked at the returned immutable
commit by LFS SHA-256 or git-blob SHA-1, and only then removed locally.  The
final `metadata/release_summary.json`, `metadata/quarantine.jsonl`, batch
manifests, and `metadata/remote_verification.json` are authoritative for final
counts and hashes.

## License and citation

HOI4D is released under CC BY-NC 4.0.  This derivative remains non-commercial.
Users must obtain MANO model assets under the MANO license themselves.  Cite
the original HOI4D paper when using this release.